DE60031432T2 - SYSTEM, METHOD, AND MANUFACTURED SUBJECT FOR DETECTING EMOTIONS IN LANGUAGE SIGNALS BY STATISTICAL ANALYSIS OF LANGUAGE SIGNAL PARAMETERS - Google Patents

Abstract

A computer system monitors a conversation between an agent and a customer. The system extracts a voice signal from the conversation and analyzes the voice signal to detect a voice characteristic of the customer. The system identifies an emotion corresponding to the voice characteristic and initiates an action based on the emotion. The action may include communicating the emotion to an emergency response team, or communicating feedback to a manager of the agent, as examples.

Description

Territory of invention

The The present invention relates to speech recognition, and more particularly on detecting emotions using statistics, which for Speech or voice signal parameters are calculated or be.

background the invention

Even though the first monograph about an expression of emotions in animals and humans by Charles Darwin was written in the last century and psychologists gradually Knowledge in the field of emotion detection and voice recognition There has been a new wave of interest recently both of psychologists as well as artificial intelligence specialists drawn. There are several reasons for this renewed interest: technological progress in recording, Storage and processing of audiovisual information; the development non-intrusive sensors; the introduction or occurrence of portable computers; and the compulsion, the human-computer interface of Point and click to a sensation and feel to expand. Farther became a new field of research in AI or artificial Intelligence as mind-related or emotional computation is known, recently identified.

Concerning Research in recognizing emotions in the language have one hand Psychologists performed many experiments and suggested theories. On the other hand, Al researchers contributed to the following areas: Synthesis from emotional language, emotion recognition and use of agents or means for decoding and expressing Emotions. A similar one Progress has been made in speech recognition.

In spite of research into recognizing emotions in language was the Technique without methods and devices that an emotion detection and use voice recognition for business purposes.

The Reference "Pattern Recognition in the Vocal Expression of Emotional Categories ", Jimenez-Fernandez et al, Proceedings of the Ninth Annual Conference of IEEE Engineering in Medicine and Biology Society, Vol. 4, 13-16, Nov. 1987, p 2090-2091, discloses the use of acoustic parameters and Statistics of such parameters when detecting an emotion a speaker. The parameters used are: variable, which refer to the "tonal contour"; the average fundamental frequency and its deviation; Record length (containing stressed, unstressed and silent parts); and average or average energy of the sentence and its variability.

The Reference "Emotion Recognition and Synthesis System on speech ", T. Moriyana et al., June 1999, p 840-844, Proc IEEE Int. Conf. on Multimedia Computing and Systems, reveals a in Reference Set statistical parameters of prosody to the emotional Content of language.

According to a first Aspect of the invention is a method for detecting emotion in a language or voice according to claim 1 available posed.

According to one second aspect of the invention is a computer program which embodied on a computer readable medium for detecting of emotion in a voice according to claim 2.

According to one Third aspect of the invention is a system for detecting Emotion in a language or voice according to claim 3 provided.

Embodiments of the invention provide a system, method, and computer program for detecting emotion using statistics. First, a database is made available. The database includes statistics including statistics of human associations of voice parameters with emotions. Next, a voice signal is received. At least one feature is extracted from the speech signal. Then the extracted speech feature is compared with the speech parameters in the database. An emotion is selected from the database based on the comparison of the extracted speech feature with the speech parameters and then spent.

In an embodiment of the present invention the feature that is extracted or extracted has a maximum Value of a fundamental frequency, a standard deviation of the fundamental frequency, a range of fundamental frequency, an average of the fundamental frequency, an average of a Bandwidth of a first formant, an average of a bandwidth a second formant, a standard deviation of an energy, a speech rate, a slope the fundamental frequency, a maximum value of the first formant, a maximum value of energy, a range of energy, a Area of the second formant and / or an area of the first Formants.

In another embodiment of the present invention the database probabilities of a particular or specific Speech feature associated with an emotion. Preferably contains selecting the emotion from the database analyzing the probabilities and a selection the most probable emotion based on the probabilities. Optional can the probabilities of database performance confusion statistics contain. Also optional, the statistics in the Database self-detection statistics included.

Short description the drawings

The The invention will be better understood when the following detailed Description taken into account becomes. Such a description makes reference to the attached drawings, wherein:

1 Fig. 12 is a schematic diagram of a hardware implementation of an embodiment of the present invention;

2 Fig. 10 is a flowchart illustrating an embodiment of the present invention which detects emotion using a voice analysis;

3 Fig. 10 is a graph showing the average accuracy of recognition for a s70 record;

4 is a map illustrating the average accuracy of recognition for an s80 record;

5 Fig. 10 is a graph illustrating the average accuracy of recognition for a s90 record;

6 Fig. 10 is a flowchart illustrating an embodiment of the present invention which detects emotion using statistics;

7 Fig. 10 is a flowchart illustrating a method for detecting nervousness in a voice in a business environment to help prevent fraud;

8th Fig. 10 is a flow chart illustrating an apparatus for detecting emotion from a voice sample in accordance with an embodiment of the present invention;

9 Fig. 10 is a flow chart illustrating an apparatus for producing visible recordings of sound in accordance with an embodiment of the invention;

10 Fig. 10 is a flow chart illustrating an embodiment of the present invention which monitors emotions in voice signals and provides feedback based on the detected emotions;

11 Fig. 10 is a flow chart illustrating a system that compares user vs. computer emotion detection of voice signals to enhance emotion recognition of either an embodiment of the invention, a user, or both;

12 is a schematic diagram in block form of a speech recognition device;

13 a schematic diagram in block form of the element arrangement and a memory block in 12 is;

14 a speech recognition system with a biomonitor and a preprocessor illustrated;

15 a biosignal illustrated by the biomonitor of 14 was generated;

16 illustrates a circuit within the biomonitor;

17 is a block diagram of the preprocessor;

18 illustrates a relationship between a pitch modification and the biosignal;

19 Fig. 10 is a flowchart of a calibration program;

20 generally shows the configuration of the portion of the system, wherein an improved selection of a set of pitch period candidates is obtained.

detailed description

In accordance with at least one embodiment of the present invention, a system for performing various functions and activities through voice analysis and voice recognition is provided. The system may be implemented using a hardware implementation, such as those described in U.S. Patent Nos. 4,766,866 1 is illustrated. Furthermore, various functional and user interface features of an embodiment of the present invention may be implemented using software programming, such as object-oriented programming (OOP).

Hardware Overview

A representative hardware environment of a preferred embodiment of the present invention is shown in FIG 1 which illustrates a typical hardware configuration of a workstation which is a central processing unit 110 , such as a microprocessor, and a number of other units that communicate over a system bus 112 are connected. In the 1 shown workstation includes a random access memory (RAM) 114 , a read-only memory (ROM) 116 , an I / O adapter 118 for connecting peripheral devices such as floppy disk storage units 120 by bus 112 , a user interface adapter 122 for connecting a keyboard 124 , a mouse 126 , a speaker 128 , a microphone 132 , and / or other user interface devices, such as a touch screen (not shown) to the bus 112 , Communication adapter 134 for connecting the workstation to a communications network (eg, a data processing network) and a display adapter 136 to connect the bus 112 with a display device 138 , The workstation typically has an operating system based thereon, such as the Microsoft Windows NT or Windows / 95 operating system (05), the IBM OS / 2 operating system, the MAC OS, or UNIX operating system.

emotion recognition

The The present invention is directed to using recognition of Emotions in the language for business purposes directed. Some embodiments of the present inven tion can used to reflect the emotion of a person based on a person To detect voice analysis and the detected emotion of the person issue. Other embodiments of the present invention for the Detecting the emotional state in telephone call center conversations and providing a feedback or a feedback for an operator or supervisor for surveillance purposes be used.

• • Daten sollten von Leuten gefordert bzw. gesammelt werden, welche nicht professionelle Schauspieler oder Schauspielerinnen sind, um die Genauigkeit zu verbessern, da Schauspieler und Schauspielerinnen eine bestimmte Sprachkomponente überbetonen könnten, was einen Fehler erzeugt.
• • Daten könnten von Testsubjekten gefordert werden, welche aus einer Gruppe ausgewählt sind, von welcher erwartet wird, daß sie analysiert wird. Dies würde die Genauigkeit verbessern.
• • Auf Sprache in Telefonqualität (< 3,4 kHz) kann abgezielt werden, um eine Genauigkeit zur Verwendung mit einem Telefonsystem zu verbessern.
• • Die Erprobung kann auf nur einem Stimmsignal beruhen. Dies bedeutet, daß die modernen Spracherkennungstechniken ausgeschlossen würden, da diese eine viel bessere Qualität des Signals und Rechenleistung erfordern.

If the target subjects are known, it is suggested that a study be run on some of the target subjects to determine which portions of a vote are most reliable as indicators of an emotion. If target subjects are not available, other subjects or persons can be used. Taking this orientation into consideration, the following discussion applies:

• • Data should be requested / collected by people who are not professional actors or actresses to improve accuracy, as actors and actresses could overemphasize a particular language component, creating an error.
• Data may be required from test subjects selected from a group expected to be analyzed. This would improve the accuracy.
• • Telephone grade (<3.4 kHz) voice can be targeted to improve accuracy for use with a telephone system.
• • The trial can be based on just one voice signal. This means that modern speech recognition techniques would be ruled out since they require much better signal quality and processing power.

Data Collection & Evaluation

• • "Dies ist nicht, was ich erwartete."
• • "Ich werde da sein."
• • "Morgen ist mein Geburtstag."
• • "Ich werde nächste Woche heiraten."

In an exemplary test, four short sentences from each of thirty people are recorded or recorded:

• • "This is not what I expected."
• • "I'll be there."
• • "Tomorrow is my birthday."
• • "I will marry next week."

Everyone Sentence should be five times to be recorded; each time the subject portrays one of the following emotional states: Happiness, anger, sadness, Anxiety / nervousness and normal (unemotional). five Subjects can also the sentences Record twice with different recording parameters. Thus, each subject has recorded 20 or 40 statements which a stock yielding 700 statements with 140 statements per emotional State contains. Each statement can be made using a near-by microphone to be recorded; the first 100 statements at 22-kHz / 8-bit and the remaining 600 statements at 22-kHz / 16-bit.

• • Wie gut können Leute ohne spezielles Training Emotionen in der Sprache porträtieren bzw. darstellen und erkennen?
• • Wie gut können Leute ihre eigene Emotionen erkennen, welche sie 6-8 Wochen früher aufzeichneten?
• • Welche Arten von Emotionen sind leichter/schwerer zu erkennen?

After the inventory has been created, an experiment can be performed to find the answers to the following questions:

• • How well can people without special training portray or portray emotions in language?
• • How well can people recognize their own emotions that they recorded 6-8 weeks earlier?
• • Which types of emotions are easier / harder to recognize?

One important result of the experiment is a selection of a sentence the most reliable Statements, i. Statements or statements, which are recognized by most people. This sentence can as Training and test data for Pattern recognition algorithms are used on a computer to run.

One interactive program of a type known in the art is, can be used to put the statements in random order select and reproduce and allow a user to make any statement according to their emotional content. For example can twenty-three subjects or persons at the evaluation level and additional 20 of those who participated earlier in the admission state to have.

Table 1 shows a performance confusion matrix resulting from data gathered from the performance of the previously discussed study. The rows and columns each represent true & evaluated categories. For example, the second series states that 11.9% of statements portrayed as happy were rated normal (unemotional), 61.4% as really happy, 10.1% as annoying, 4.1% as sad , and 12.5 as anxious. It is also seen that the most obvious category is anger (72.2%) and the least recognizable category is anxiety (49.5%). A lot of confusion is found between sadness and anxiety, sadness and unemotional state, and happiness and anxiety. The mean accuracy is 63.5%, which agrees with the results of the other experimental studies. Table 1 Power Confusion Matrix

Table 2 shows statistics for evaluators for each emotional category and for a combined performance, which was calculated as the sum of performances for each category. It can be seen that the variance for anger and sadness is much less than for the other emotional or emotional categories. Table 2 Statistics of the evaluators

Table three below shows statistics for "actors," ie, how well subjects portray emotions. More specifically, the numbers in the table show which portion of portrayed emotions of a particular category was recognized as this category by other subjects. It is interesting to see that in a comparison of Tables 2 and 3, the ability to portray emotions (overall mean is 62.9%) remains at approximately the same level as the ability to detect emotions (overall mean is 63.2%), however, the variance is much larger for portraying. Table 3 Statistics of the actors

Table 4 shows self-referential statistics, ie how well subjects were able to recognize their own portraits. We can see that people are much better at recognizing their own emotions (the mean is 80.0%), especially for anger (98.1%), sadness (80.0%) and anxiety (78.8%). Interestingly, anxiety was better recognized than happiness or happiness. Some subjects failed to recognize their own portrayals of happiness and normality. Table 4 Self-referral statistics

From the inventory of 700 statements, five nested records containing utterances recognized as representing the given emotion by at least p percent of the subjects (p = 70, 80, 90, 95, and 100%) can be selected , For the current discussion, these records should be referred to as s70, s80, s90, s95 and s100. Table 5 below shows the number of elements in each record. We can see that only 7.9% of the utterances of the stock were recognized by all subjects. And this number increases in a straight line up to 52, 7% for the record s70, which agrees with the 70% level of concordance in decoding emotions in speech. Table 5 p-level match records

The Results provide a useful Insight about human performance and can be used as a baseline for one Serve comparison with a computer presentation or performance.

feature extraction

• • Stimmenergie bzw. Vokalenergie
• • spektrale Frequenz-Merkmale
• • Formanten (üblicherweise werden nur ein oder zwei erste Formanten (F1, F2) betrachtet).
• • zeitliche Merkmale (Sprachtempo und Unterbrechung).

It has been found that the pitch of the main voice notes is for emotion recognition. Strictly speaking, the pitch is represented by the fundamental frequency (F0), ie the main (lowest) frequency of the vibration or vibration of the vocal folds or vocal cords. The other acoustic variables that contribute to vocal emotion signaling are:

• • Voice energy or vocal energy
• • Spectral frequency characteristics
• • Formants (usually only one or two first formants (F1, F2) are considered).
• • temporal characteristics (speed and interruption).

A other approach to a feature extraction is the set of features by considering some derivative features, such as LPC (linear, predictive, coding parameters) of a signal or features of the smoothed pitch contour and their derivatives.

For this Invention, the following strategy can be applied. First, consider the fundamental frequency F0 (i.e., the main (lowest) frequency of the Vibration of the vocal cords), Energy, speech speed, the first three formants (F1, F2 and F3) and their bandwidths (BW1, BW2 and BW3) and calculate for them as many statistics or statistical data as possible. Then rank the statistics using feature selection techniques, and choose a set of "most important" features.

The Speech speed can be considered the inverse of the average length of the voiced part of an utterance become. For all other parameters can the following statistical data are calculated: mean, Standard deviation, minimum, maximum and range. In addition, can for F0 the slope as a linear regression for the voiced part of the Language can be calculated, i. that line leading to the pitch contour fits. The relative voiced or voiced energy can also be called the proportion of the voiced energy to the total energy of the utterance is calculated become. In total, there are about 40 features for each utterance.

Of the RELIEF-F or FULL-F algorithm can be used for feature selection become. For example, ENTLASTEF can run on the s70 record be the number of the next Neighbor varies from 1 to 12, and the characteristics accordingly their sum of ranks or rows are ordered. The top 14 features are the following: F0 maximum, F0 standard deviation, F0 range, F0 average, BW1 mean, BW2 mean, energy standard deviation, speech speed, F0 Slope, F1 maximum, energy maximum, energy range, F2 range and F1 range.

Around to investigate how sentences of features affect the accuracy of emotion recognition algorithms, can three nested sentences formed of features based on their sum of rankings become. The first sentence contains the top eight features (from F0 maximum speech rate), the second sentence extends from the first to two more features (F0 slope and F1 maximum), and the third set contains all 14 maximum features. More details about the RELIEF-F algorithm is used in the publication Proc. European Conf. On Machine Learning (1994) in the article by I. Kononenko, entitled "Estimation Attributes: Analysis and extension of "RELIEF" or "RELIEF" set out and on found on pages 171-182.

2 illustrates an embodiment of the present invention that detects emotion using voice analysis. In act 200, a vocal signal is received, such as by a microphone or in the form of a digitized sample. A predetermined number of features of the voice signal are extracted as set forth above and in the process 202 selected. These features include, but are not limited to, a maximum value of a fundamental frequency, a standard deviation of the fundamental frequency, a range of the fundamental frequency, an average of the fundamental frequency, an average of a bandwidth of a first formant, an average of a bandwidth of a second formant, a Standard deviation of energy, a speech rate, a slope of the fundamental frequency, a maximum value of the first formants, a maximum value of the energy, a range of energy, a range of the second formants, and a range of the first formants. Using the in function or process 202 selected features, an emotion that is associated with the voice signal, in operation 204 determined based on the extracted feature. Finally, in process 206 the particular emotion spent. See the discussion below, especially with reference to 8th and 9 for a more detailed discussion of determining an emotion based on a vocal signal in accordance with the present invention.

Preferably For example, the feature of the voice signal is selected from the group of features which from the maximum value of the fundamental frequency, the standard deviation of Fundamental frequency, the range of the fundamental frequency, the mean of the Fundamental frequency, the mean of the bandwidth of the first formant, the mean of the second formant bandwidth, the standard deviation the energy, and the speech rate exist. Ideally the extracted feature includes at least one of the slopes the fundamental frequency and the maximum value of the first formants.

optional is a variety of features extracted, including the maximum value of the fundamental frequency, the standard deviation of the fundamental frequency, the range of the fundamental frequency, the mean of the fundamental frequency, the mean value of the bandwidth of the first formant, the mean the bandwidth of the second formant, the standard deviation of the Energy, and the speech rate. Preferably include the extracted features the slope of the fundamental frequency and the Maximum value of the first formants.

When another option will extract a variety of features including the maximum value of the fundamental frequency, the standard deviation the fundamental frequency, the range of the fundamental frequency, the mean of the Fundamental frequency, the mean of the bandwidth of the first formant, the mean of the bandwidth of the second formant, the standard deviation the energy, the speech rate, the slope of the fundamental frequency, the maximum value of the first formant, the maximum value of the energy, the area of energy, the area of the second formant, and the area of the first formant.

computer performance

Around Recognizing emotions in one language can be two exemplary approaches neural networks and ensembles of sorters or classifying machines. In the first approximation can be a two-ply spreading backwards neural network architecture with an 8, 10 or 14 element input vector, 10 or 20 knots in the hidden sigmoid layer and five knots be used in the output linear layer. The number Spending matches the number of emotional categories. To train and test the algorithms, records s70, s80 and s90 are used. These sentences can happen in training (67% of utterances) and test (33%) subsets split or split. Various neural network classifiers, which trains with different starting weight matrices are, can be created or generated. This approach when connected to the s70 record and the 8-feature set applied above, gave the average accuracy of about 55% with the following distribution for emotional categories. normal state is 40-50%, cheerfulness is 55-65%, trouble is 60-80%, sadness is 60-70%, and anxiety is 20-40%.

For the second approach or second access ensembles of classifiers are used. An ensemble consists of an odd number of neural network classifiers, which on different subsets or subsets of Training set using bootstrap aggregation and crossover or cross-confirmed Committee techniques were trained. The ensemble falls Decisions based on the majority voting principle proposed or recommended ensemble sizes range from 7 to 15.

3 shows the average accuracy of detection for a s70 data set, all three sets of features, and both neural network architectures (10 and 20 neurons in the hidden layer). It can be seen that the accuracy for happiness remains the same (approximately ~ 68%) for the different sets of features and architectures. The accuracy for anxiety is pretty low (15-25%). Aggravation accuracy is relatively low (40-45%) for the 8-feature set and improves dramatically (65%) for the 14-feature set. However, the accuracy for sadness is higher for the 8-feature set than for the other sentences. The average accuracy is about 55%. The low accuracy for fear confirms the theoretical result, which states that if the individual classifiers make uncorrelated errors at rates that exceed 0.5 (it is 0.6-0.8 in our case), then the error rate of the chosen one Ensembles increases.

4 shows results for a s80 record. It is seen that the accuracy for the normal state is low (20-30%). The accuracy for anxiety changes dramatically from 11% for the 8-feature set and 10-neuron architecture to 53% for the 10-feature and 10-neuron architecture. The accuracy for happiness, anger and sadness is relatively high (68-83%). The average accuracy (~ 61%) is higher than for the s70 data set.

5 shows results for a s90 record. We can see that anxiety accuracy is higher (25-60%) but follows the same pattern as shown for the s80 record. Accuracy for sadness and anger is very high: 75-100% for anger and 88-93% for grief. The average accuracy (62%) is approximately equal to the average accuracy for the s80 data set.

6 illustrates an embodiment of the present invention that detects emotion using statistics. First, a database is in process 600 made available. The database includes statistics that include statistics of human associations of voice parameters with emotions, such as those shown in the tables above and 3 to 5 are shown or are. Furthermore, the database may include a series of pitches that are associated with anxiety and another series of pitches associated with happiness and an error range for particular pitches. Next, a voice signal becomes in action 602 received or recorded. In process 604 one or more feature (s) is extracted from the vocal signal. See the feature extraction section above for more details on extracting features from a vocal signal. Then in process 606 compared the extracted voice memo with the voice parameters in the database. In process 608 For example, an emotion is selected from the database based on the comparison of the extracted voice feature with the voice parameters. This may include, for example, comparing digitized speech samples from the database with a digitized sample of the feature extracted from the vocal signal to produce a list of probable emotions and then using algorithms to obtain human accuracy statistics in recognizing the emotion to make a final determination of the most likely emotion. The selected emotion becomes final in action 610 output. See the section titled "Exemplary Devices for Detecting Emotion in Speech Signals," below, for computerized mechanisms to perform emotion recognition in a language.

In One aspect of the present invention includes the database Probabilities of special characteristics, which with a Emotion are associated. Preferably, the selection involves the emotion from the database analyzing the probabilities and a Choose the most likely emotion based on the probabilities. Optionally the probabilities of database performance confusion statistics include, for example, in the editing confusion matrix shown above. Also optional, the statistics in the Database include self-detection statistics, such as those shown in the tables above.

In In another aspect of the present invention, the feature includes which is extracted, a maximum value of a fundamental frequency, a standard deviation of the fundamental frequency, a range of the fundamental frequency, an average of the fundamental frequency, an average of a bandwidth a first formant, an average of a bandwidth of a second formant, a standard deviation of energy, a speech rate rate, a slope of the fundamental frequency, a maximum value of first formant, a maximum value of energy, an area of Energy, an area of the second formant, and / or an area the first formant.

7 Figure 10 is a flow chart illustrating a method for detecting nervousness in a voice in a business environment to help prevent fraud. First in process 700 Receive voice signals from a person during a business event. For example, the voice signals may be generated by a microphone in the vicinity of the person, may be captured by a telephone tap, etc. The voice signals will be in action during the business event 702 analyzed to determine a level of nervousness of the person. The voice signals can be analyzed as set forth above. In process 704 an indication is given of the level of nervousness, preferably before the business event is completed, so that anyone attempting to prevent fraud may make an assessment of whether the person is to confront before it Person goes away. Any type of output is acceptable, including a paper printout or display on a computer screen. It should be understood that this embodiment of the invention can detect emotions other than nervousness. Such emotions include stress and any other emotion specific to a person when committing a fraud.

These embodiment of the present invention has a particular application in business areas, such as contract negotiations, insurance settlements, Customer service, etc. Fraud in these areas costs companies millions every year. Fortunately the present invention provides a tool to in the fight to help such fraud. It should also be noted that the present invention Law enforcement applications, such as also in a courtroom environment etc.

Preferably, a degree of certainty regarding the person's level of nervousness is given to help someone in a search for fraud make a determination as to whether the person has fraudulently spoken. This may be based on statistics as described above in the embodiment of the present invention with reference to FIG 6 was set out. Optionally, the person's level of nervousness can be displayed in real time to allow someone attempting to prevent fraud to obtain results very quickly so that he or she is able to challenge the person soon after the person becomes one makes a suspicious statement.

When another option, the level of nervousness display may include an alarm, which triggered will if the level of nervousness exceeds a predetermined level. The alarm can be a visual notification on a computer display include, an audible Sound, etc., to alert a supervisor, the listener, and / or someone looking for fraud. The alarm could also be with a recording device which would begin to record the conversation when the alarm is triggered if the conversation is not already recorded.

The Alarm options would be especially useful in a situation be where many people take turns talking. An example would be in a customer service department or on the phone of a customer service representative. As every customer's turn comes to deal with a customer service representative to speak the present invention the level of nervousness in the language of Detect customers. If the alarm was triggered because of the level of nervousness a customer crossed the prescribed level, the Customer service representative or manager through a visual or visible display on his or her computer screen, a flashing light, etc. are notified. The customer service representative, now of the possible Knowing cheating could then try to expose the fraud if he exist. The alarm could also be used to notify a manager as well. About that out could start a recording of conversation or conversation after the alarm has been activated.

In an embodiment In the present invention, at least one feature of the voice signals is extracted and used to the nervous level to determine the person. Features that are extracted may include a maximum value of a fundamental frequency, a standard deviation of Fundamental frequency, a range of the fundamental frequency, an average value the fundamental frequency, an average of a bandwidth of a first Formants, an average of a bandwidth of a second formant, a standard deviation of energy, a speech rate, a slope or slope of the fundamental frequency, a maximum value the first formant, a maximum value of energy, an area energy, an area of the second formant, and an area the first formant. Thus, for example, a degree of fluctuation in the tone of the voice, as he from measurements or readings of the fundamental frequency was determined to be used in determining a level of nervousness or levels to help. The greater the degree a waver, the higher is the level of nervousness. Breaks in the language of the person can also be considered become.

Of the The following section describes devices that are used can, to include emotion, nervousness, in voice signals.

Exemplary devices for detecting emotion in voice or voice signals

This Section describes various devices for analyzing of language in accordance with the present invention.

An embodiment of the present invention includes a device for analyzing a person's speech to determine their emotional state. The analyzer operates on the real-time frequency or pitch components within the first formant band of human speech. In analyzing the speech, the device analyzes appearance patterns of particular values with respect to patterns of differential first formant pitch, rate of pitch change, duration and time interval development. These factors relate in a complex but very fundamental way to both transient and emotional long-term states.

The Human speech is through two basic sound generation mechanisms stimulated. The vocal cords; thin stretched or stretched membranes under muscle control or oscillate or swing when ejected Air from the lungs passes through them. They produce a characteristic "hum" sound at one fundamental frequency between 80 Hz and 240 Hz. This frequency will over a moderate range by both conscious and unconscious muscle contraction and relaxation varies. The waveform of the basic "humming" includes many Harmonics, some of which differ in excitation resonance is fixed and changeable cavities associated with the vocal area. The second keynote, which during a language is a pseudorandom noise, which is a has fairly broad and uniform frequency distribution. It is caused by turbulence as soon as exhaled or expelled air moves through the vocal tract and is called a "hissing" sound or sibilant. He will mainly modulated by tongue movements and also excites the specified and changeable Cavities. It is this complex mix of "humming" and "hissing" sounds that are shaped by the resonant cavities and articulate which language to produce.

In an energy distribution analysis of speech sounds be found that the Energy in separate or distinct frequency bands, formants called, falls. There are three significant formants. The system described here uses the first formant band, which ranges from the basic "buzz" frequency to about 1000 Hz stretches. This band not only has the highest energy content, but reflects a high degree of frequency modulation as a function of various vocal tract and facial muscle tension variations.

In Reality becomes by analyzing certain distribution patterns the first formant frequency is a qualitative measurement of speech-related Muscle tension variations and interactions performed. There this muscle is predominantly through secondary unconscious operations be biased and articulated, which in turn by a affects emotional state are, a relative measure of an emotional activity can be independent of the consciousness a person or a lack of consciousness determined by this condition become. The research confirms also a general conjecture that, since the mechanisms of language extremely complex and are largely autonomous, very few people are capable of consciously one to "project" fictional emotional state. Actually generated an attempt to do that, usually his own unique psychological stress "fingerprint" in the vocal pattern.

Because of the characteristics of the first formant speech sounds or -Signals analyzed the present invention is an FM demodulated first formant speech signal and generates an output indicating zeros thereof.

The Frequency or number of zeros or "flat" points in the FM demodulated signal, the length the zeros and the ratio the total time at which zeros exist during a word period, for Total time of the word period are all for an emotional state indicative of the individual or indicative. By looking at the Issue of the device, the user or the user's appearance to see or feel the zeros and thus by observing the output of the number or frequency of zeros determine the length zeros and ratio the total time while which zeros during a word period up to the length the word period exist, the emotional state of the individual or individual.

In the present invention, the first formant frequency bath of a speech signal is FM demodulated and the FM demodulated signal is applied to a word detector circuit which detects the presence of an FM demodulated signal. The FM demodulated signal is also applied to zero detector means which detect the zeros in the FM demodulated signal and produce an output indicative thereof. An output circuit is coupled to the word detector and the zero detector. The output circuit is enabled by the word detector when the word detector detects the presence of an FM demodulated signal, and the output circuit generates an output indicative of the presence or absence of a zero in the FM demodulated signal. The output of the output circuit is displayed in a manner in which it is perceived by a user, so that the user is provided with an indication of the existence of zeroes in the FM demodulated signal. The user of the device thus monitors the zeros and thereby can determine the emotional state of the individual, their language is analyzed.

In another embodiment In the present invention, the vocal vibrato is analyzed. The so-called Voice vibrato was created as a semi-voluntary response, which while studying a misdirection could be of value along with certain other reactions; as for example, respiratory volume; Inspiratory expiratory ratios; Metabolic rate; regularity and rate of inhalation; Association of words and ideas; Facial expressions; Motor response; and reactions to certain narcotics; however No suitable technique has been previously developed which is a valid and reliable Analysis of voice changes in the clinical determination of an emotional state, opinions or deception attempts of a subject or a person.

Early experiments, which included attempts to change voice quality with emotional stimuli to correlate, have established that the human Language is affected by strong emotion. Detectable changes the voice appears much more quickly, following a stress stimulation, as it is the classic indications of physiological manifestations do, which result from the functioning of the autonomic nervous system.

Two Change types of a voice as a result of stress. The first of these is called the coarse change, which is usually only as a result of a significant stress situation. This change manifests itself in audible noticeable changes the speech rate, volume, voice shake, change in the space between syllables, and a change in pitch or frequency the voice. This rough change is the subject of the conscious Control, at least in some subjects or persons, if the stress level below that of a total loss of control.

Of the second type of voice change is that of voice quality. This type of change is not for that human ear distinguishable or perceptible, but is one apparently unconscious Manifestation of the slight tension of the vocal cords under even slight stress, resulting in a damping selected Frequency variations. When graphing the difference easy between unstressed or normal vocalization and vocalization under mild or low Stress, Deception attempts or hostile or opposing attitudes are perceptible. These patterns have over a wide range of human voices of both sexes, different Age and under different situational conditions as true or correctly proved. This second change type is not subject a conscious one Control.

It are two types or types of sound or sound, which by the human vocal anatomy are generated. The first type of sound is a product of the vibration of the vocal cords, which in turn is a product a partial closing of the glottis is and a forcing of air through the glottis by a contraction of the lung cavity and lungs. The Frequencies of these vibrations can generally between 100 and 300 hertz, depending on gender and age of the speaker and intonations the speaker uses. This tone has a rapid decay or cooldown.

The second type of sound includes the formant frequencies. These form a sound resulting from the resonance of the cavities in the head, involving the neck, mouth, nose and sinuses. This sound is through an excitation of the resonant cavities through a sound source of lower frequencies, in the case of by the vocal cords produced vocalized sound, or by a partial restriction of the Passage of air from the lungs, as in the case of unvoiced Generated friction sounds. Whatever the source of excitation, the Frequency of formants is affected by the resonance frequencies of the Cavity determined. The formant frequencies generally appear at about 800 hertz and appear in certain frequency bands, which coincide with the resonant frequency of the individual cavities. The first, or lowest, formant, is those formed by the mouth and throat or throat cavities is and is noticeable for their frequency shift once the mouth is its dimensions and Volume at the formation of different sounds, especially of vowels changes. The highest Formant frequencies are more constant due to the more constant volume the cavities. The formant waveforms are ringing signals, in contrast to the rapidly decaying vocal cord signals. If voiced sounds or sounds be voiced the voice waveforms are referred to the formant waveforms as Impressed amplitude modulations or superimposed.

It has been discovered that a third signal category exists in the human voice and that this third signal category is related to the second type of voice change, as discussed above. This is an infrasonic or subsonic frequency modulation, which is present to some degree in both vocal cord sounds and formant sounds. This signal is typically between 8 and 12 hertz. Accordingly, it is not audible to the human ear. Because of the fact that this characteristic forms a frequency modulation that differs from amplitude modulation, it is not directly perceptible on time base / amplitude map records. Because of the fact that this infrasonic signal is one of the more significant vocalizations of psychological stress, it will be treated in greater detail.

It There are various analogies which are used to describe schematic Presentations of the entire voting process available put. Both mechanical and electronic analogies will be used successfully for example in the design of computer voices or applied. However, these analogies look at the voiced Sound source (vocal cords) and the walls the cavities as hard and constant features. However, both the vocal cords pose as also the walls the basic formant-generating cavities in reality a flexible one Tissue, which is instantaneous on the complex arrangement of Muscles providing a control of the tissue will respond. These Muscle, which is the vocal cords controlled by the mechanical connection of bone and cartilage both the targeted and automatic generation of vocal sound and change the voice pitch by an individual. In similar Way these muscles allow the tongue, lips and neck respectively Throat control, both targeted and automatic control the first formant frequencies. Other formants may be similar to be influenced to a more limited degree or extent.

It is worth to note that while normal Language these muscles on a small percentage of their total workability work. For this reason, despite being used to change position, they remain the vocal cords and the positions of the lips, tongue and inner neck walls, the Muscles in a relatively relaxed state. It was determined that during this relatively relaxed state a muscle wave movement typically at the aforementioned Frequency of 8-12 hertz occurs. This wave motion causes a slight variation in the tension of the vocal cords and causes shifts of the basic pitch frequency the voice. Also, the waveform slightly varies the volume of the Resonant cavity (especially that associated with the first formant) and the elasticity of the Cavity walls, to cause shifts in the formant frequencies. These Shifts about a central frequency form a frequency modulation the central or carrier frequency.

It is important to note that no the shifts in the basic pitch frequency of the voice or in the formant frequencies directly detectable by a listener is, partly because the shifts are very small, and partly, because they are primary in the aforementioned inaudible Frequency range exist.

Around observing this frequency modulation can be any of several existing techniques for demodulating a frequency modulation be used, of course, taking into account that the modulation frequency the nominal 8-12 hertz and the. Wear one of the tapes inside the voice spectrum is.

Around complete To understand the above discussion must have the concept of a "center of gravity" of this waveform be understood. It is possible, approximately the midpoint between the two extremes or extremes of each individual deflection of the recording pin to determine. When the midpoints are marked between the extremes of all the deflections and if these centers are then roughly through a continuous Curve, it will be seen that a line that is one Average or "center of gravity" of the entire waveform approaches, will result. A combination of all such markings, with some smoothing, results in a gentle curved Line. The line represents the infrasound frequency modulation, which from the previously described Waveforms results.

As mentioned above, was determined that the Arrangement of with the vocal cords Associated muscles and cavity walls subject to a gentle Muscle tension is when mild to moderate psychological Stress in the individual review or Examination is generated. This tension, for the subject or the person is imperceptible, and similar through normal, unsupported Observation techniques for not the reviser is noticeable, is sufficient to lower the muscle waveforms or virtually reduce or almost eliminate muscle waveforms, which in the non-stressed subject which removes the basis for the carrier frequency variations which produce the infrasound frequency modulations.

While the Use of infrasound waveform unique to the technique is which voice as the physiological medium to the physiological stress evaluation uses, the voice provides additional instrumented displays of non-audible physiological changes as a result of psychological stress, which undergoes similar physiological changes Detect techniques and devices in current use are. Of the four most common used physiological changes, which previously mentioned (brain wave pattern, heart activity, skin conductivity and Breathability) affect two of these, the respiratory activity and cardiac activity, directly and indirectly the amplitude and detail of a waveform of a oral statement and provide the basis for a coarser one Evaluation of psychological stress, especially if the test or testing involves sequential vocal responses, to disposal.

Another device is in 8th shown. As shown, a converter converts 800 The sound waves of the subject's verbal utterances are converted into electrical signals, with the input of an audio amplifier 802 which is simply for the purpose of enhancing the performance of the electrical signals to a more stable, more useful level. The output or output of the amplifier 802 is with a filter 804 which serves primarily for the purpose of removing some unwanted low frequency components and noise components.

After filtering, the signal is sent with an FM discriminator 806 connected, wherein the frequency deviations are converted from the central or center frequency into signals which vary in amplitude. The amplitude varying signals are then in a detector circuit 808 for the purpose of rectifying the signal and generate a signal which forms a series of half-wave pulses. After detection, the signal becomes an integrator circuit 810 in which the signal is integrated to the desired extent. In the circuit or in the circuit 810 the signal is either integrated to a very small extent, generating a waveform, or integrated to a greater extent, generating a signal. After integration, the signal in the amplifier 812 amplified and with a processor 814 which determines the emotion associated with the voice signal. An output device 816 , such as a computer screen or printer is used to output the detected emotion. Optionally, statistical data can also be output.

A somewhat simpler embodiment of a visual record generating apparatus in accordance with the invention is shown in FIG 9 shown, wherein the acoustic signals through a microphone 900 are converted into electrical signals which magnetically in a tape recorder 902 to be recorded. The signals can then be processed by the remaining equipment at various speeds and at any time, with reproduction using a conventional semiconductor diode 904 connected, which rectifies the signals. The rectified signals are input with a conventional amplifier 906 connected and also with the movable contact of a selector switch, which in general 908 is displayed or designated. The movable contact of the switch 908 can be moved to any one of a plurality of fixed contacts, each of which is connected to a capacitor. In 9 is a selection of four capacitors 910 . 912 . 914 and 916 each having one terminal connected to a predetermined contact of the switch and the other terminal connected to ground. The output or output of the amplifier 906 is with a processor 918 connected.

A tape recorder which can be used in this particular set of equipment was a Uher Model 4000 four-speed tape unit having its own internal amplifier. The values of the capacitors 910 - 916 were each 0.5, 3, 10 and 50 microfarads and the input impedance of the amplifier 906 was about 10,000 ohms. As will be appreciated, various other components could be used in this device.

In the operation of the circuit of 9 becomes the by diode 904 integrating the resulting rectified waveform to the desired extent, the time constant being selected such that the effect of the frequency modulated infrasonic wave appears as a slowly varying DC level approximately following the line representing the "centroid" of the waveform. The deflections shown in this particular diagram are relatively fast, indicating that the switch was connected to one of the lower value capacitors. In this embodiment, mixed filtering is performed by the capacitor 910 . 912 . 914 or 916 , and achieved in the case of a reduction of the playback speed by the tape recorder.

telephone Operation with operator feedback

10 illustrates an embodiment of the present invention that monitors emotions in voice signals and provides operator feedback based on the detected emotions. First, a voice signal representative of a component of a conversation between at least two subjects is in operation 1000 receive. In process 1002 An emotion associated with the voice signal is determined. Finally, in process 1004 a feedback is provided to a third party based on the particular emotion.

The Conversation can be over a telecommunications network, as well as a wide area network, such as the internet when using internet telephony is used. As an option, the emotions are rested or sieved and a feedback will only be available when it is determined by the emotion, a negative emotion to be selected from the group of negative emotions, which is anger, sadness and fear exist. The same could be done with groups of positive or neutral emotions. The emotion can be through extracting a feature from the voice signal are determined as previously described in detail.

The The present invention is particularly associated with operation with an emergency response system, such as the 911 system suitable. In such a system, incoming calls could be through embodiments of the present invention become. An emotion of the caller would be during the conversation of the Caller with the technician answering the call. The emotion could then over For example, radio waves may be sent to the emergency team, i. Police, fire and / or ambulance staff, so that these over the emotional state of the caller.

In In another scenario, one of the subjects is a customer, another The subject is an employee, such as someone who hired by a call center or a customer service department is, and the third or third person is a manager. The present invention would monitor the conversation between the customer and the employee, to determine if the customer and / or the employee, for example being (becoming) excited. When negative emotions are detected, will be a feedback sent to the manager to assess the situation and intervene if necessary or can intervene.

Improve an emotion recognition

11 FIG. 12 illustrates a system that compares user and computer emotion detection of voice signals to enhance a voice recognition of one of the embodiments of the invention, a user, or both. First, in process 1100 , a voice signal and an emotion associated with the voice signal are provided. The emotion associated with the voice signal automatically goes into action 1102 determined in a manner as set forth above. The automatically determined emotion is in process 1104 stored for example on a computer readable medium. In process 1106 A user-defined emotion associated with the voice signal, which is determined by a user, is received. The automatically determined emotion comes into action with the user-defined emotion 1108 compared.

The Voice signal can be sent or issued or received by a system embodying the present invention. optional the emotion associated with the voice signal is identified, when or after the emotion is provided. In such a Case should be determined if the automatically determined emotion or the user-determined emotion matches the identified emotion. The user or user can be awarded a prize if the user-defined emotion matches the identified emotion or matches. Furthermore, the emotion can be automatically extracted by extracting at least one feature of the voice signals, such as in as discussed above.

Around can assist a user in recognizing an emotion an emotion recognition game to be played. The game could do it allow a user against a computer or another Person to attend, to see who best emotions in recorded Recognize language. A practical application of the game is autistic people in developing a better emotional Experience in helping to recognize emotion in the language.

In In one arrangement, a device may be used to transmit data via voice signals which can be used to provide emotion recognition to improve. The device accepts vocal sound a transducer such as a microphone or a sound recorder. The physical Sound wave or sound wave, which has been converted into electrical signals, runs parallel to a typical, commercially available bank used by electronic filters which the audible or Cover audio frequency range. A setting of the central or Center frequency of the lowest filter to any value, which the electric energy presentation the voice signal amplitude passes, which contains the lowest voice frequency signal, builds the center values all subsequent filters to the last, which in general the energy between 8 kHz to 16 kHz or between 10 kHz and 20 kHz, and also determines the exact number of such filters. The specific value the center frequency of the first filter is not significant as long as the deepest sounds of the human voice, about 70 Hz. Essentially any commercially available bank is applicable if to any commercially available digitizer and then microcomputer can be connected. The description section describes a specific set of center frequencies and microprocessor in the preferred embodiment. The filter quality is also not very significant, as one in the description revealed refinement algorithm each Set of filters of average quality in acceptable frequency and Amplitude values brings. The relationship 1/3 of course defines the bandwidth of all filters as soon as the center frequencies are calculated are.

this Following the sequencing process The filter output voltages are provided by a commercially available Set of digitizers or preferably multiplexers and digitizers or a digitizer digitized, which is in the same identified commercially available filter bank is installed to remove a coupling logic and hardware. Again, the quality a digitizer or a digitizer with respect to not on the speed of conversion or discrimination significant, as average currently available commercial units have the required requirements due to a correction algorithm (see specifications) and the low exceed necessary sampling rate.

Everyone complex sound or sound or sound, which is constantly changing Carries information, can with a reduction of information bits by a catch the frequency and amplitude of peaks of the signal are approximated. This is natural old knowledge, how to do it of such a process of speech signals as well. However, they were in linguistics different specific areas where such Tips often occur, referred to as "formant" regions. However, these range approaches do not always coincide with those Tips of each speaker under all circumstances. Linguist and the inventive prior art tend to be a great effort to measure "legitimized" peaks and as such, which are within the typical formant frequency ranges fall as if their definition did not involve estimates, but rather absolute or absolute values. This has numerous research and formant measuring devices causes artificially to exclude corresponding peaks that are necessary to be adequate a complex, highly variable Represent sound wave in real time. Since the present disclosure is designed to be useful for animal voice sounds even all human languages are suitable to be artificial restrictions such as formants, not of interest, and the sound or sound wave is called a complex, varying sound wave which analyzes each such sound can.

Around a peak identification regardless of a deviation in Filter bandwidth, quality normalize and simplify digitization discrimination, are the actual ones Values which for Amplitude and frequency are stored, "representative Values. "This is why so that the width of upper frequency filter is numerically similar to Bandwidth of the filter is lower frequency. Become a filter simply given consecutive values from 1 to 25, and a sound of soft to loud will be from 1 to 40 for ease on a CRT screen scaled. A correction to the frequency representation values is achieved or carried out, by changing the number of filters to a higher decimal value to the nearest integer value is set when the filter output to the right of the tip filter a larger amplitude as the filter output has to the left of the peak filter. The Details of this algorithm are in the descriptions of this disclosure described. This correction process must be done before the compression process happen while all filter amplitude values available are.

Rather than slowing the sampling rate, the preferred arrangement stores all the filter amplitude values for 10 to 15 samples per second for a speech sample of about 10 to 15 seconds prior to this correction and compression process. If the computer memory space is more critical than the throughput speed, the corrections and compression should occur between each pass, to wipe out the next one for a big strong data store. Since most commercially available average price minicomputers have sufficient memory, the preferred arrangement disclosed herein stores all the data and subsequently processes the data.

The most animal sound signals of interest, including human, include a largest amplitude peak, probably not at each end of the frequency domain. These Tip can by any simple and usual numerical sorting algorithm be determined as is done in this invention. The for amplitude and frequency representative Values are then arranged in the number three out of six memory location sets, around the amplitudes and frequencies of six peaks hold.

The highest Frequency peak over 8 kHz is placed in memory location number six and as a high frequency peak characterized. The lowest peak is in the first sentence of Storage locations or storage locations arranged. The other three are made of tips between them selected. Following this compression function, the vocal signal is replaced by a for amplitude and frequency representative Value of each of the six peaks plus a total energy amplitude from the unfiltered total signal for example ten times per Second for represents a sample or sample of ten seconds. This provides a Total number of 1300 values.

The Algorithms allow for variations in scan length in case the operator the sample length switch override with the override switch or except Strength continues to be a continuation during an unexpected noise interruption to prevent. The algorithms do this by using Averages, which are not significantly sensitive to changes in the number of samples over four or five Seconds of a sound or sound signal. The reason for a larger speech sample, if possible, is to the average "style" of the language of the Spokesman, which typically within 10 to Becomes evident for 15 seconds.

The Output of this compression function will be in the element array and fed a memory algorithm which assembles (a) four voice quality values, which are to be described below; (b) a sound "pause" or on-to-off ratio; (c) "variability" - the difference between the peak of each amplitude for the current one Pass and those of the last pass; differences between the frequency number of each peak for the current pass and that of the last run; and difference between the unfiltered Total energy of the present or current run and that the last run; (d) a "syllable change approach" by obtaining the ratio of painting that yourself the second peak more than 0.4 between runs to the total number of runs with sound changes; and (e) "high-frequency analysis" - the ratio of the number of sound-on-runs that a nonzero value in this peak for the peak amplitude number six include. This is a total of 20 elements that available per run. These are then passed to the dimension assembly algorithm.

The four voice quality values, which are used as elements are (1) the "spread" - the sample mean of all differences of runs between their average of the values representing a frequency above the maximum amplitude peak and the average of those below, (2) the "balance" - the sample mean of all average amplitude values of the runs of peaks 4, 5 & 6 divided by the average of the peaks 1 & 2. (3) "envelope level high" - the sample average of all means of runs from their amplitudes above the largest peak, divided by the biggest peak, (4) "Envelope flatness low "- the sample average from all the averages of the runs from their amplitudes below the biggest peak, divided by the largest peak.

The Voice Style Dimensions are and will be called "Resonance" and "Quality" assembled by an algorithm which defines a coefficient matrix includes on selected Elements works.

The "language style" dimensions are called "variability monotone", "choppy-soft", "staccato hold", "gently rise", "affectivity control". These five dimensions, where names belong to each end of each dimension measured and assembled by an algorithm which has a Coefficient matrix involved on 15 of the 20 sound elements detailed in Table 6 and the specification section are.

The perceptual dimension becomes "eco-structure", "invariant sensitivity", "different self "," sensory-internal "," hate-love "," independence-dependence "and" emotional-physical. "These seven perceptual dimensions with names referring to the end-of-dimensions are measured and assembled by an algorithm. which deals with a coefficient matrix and works on selected sound elements of voice and speech (detailed in Table 7) and the specification section.

A commercially available, typical computer keyboard or membrane keyboard allows the user of the present disclosure, any and all coefficients for redefinition from any compound language, voice or perception dimension for research purposes amend. Selector switches allow any or all elements or dimension values for one To display a voice sample of a given subject. The digital one Processor controls or controls the analog-to-digital conversion of the sound signal and regulates or controls the reassembly or reassembly of the vocal sound elements into numerical values of the voice, voice and perceptual dimensions.

Of the Microcomputer also coordinates the keystrokes of the operator or actuator and the chosen one Output display of values, and coefficient matrix selection to use with to co-operate with the algorithms that govern the voice, language, and perceptual dimensions put together. The output selector simply directs the output to any or all output plugs suitable for the signal to typical commercially available Monitors, modems, printers or given to a light-emitting to be addressed on-board.

By developing group profile standards using these Invention, a researcher results in publications through occupations or occupations, Malfunctions, tasks, hobby interests, cultures, languages, List gender, age, species, etc. Or the user (s) in) can compare his / her values with those of others released or with those installed in the machine.

Referring now to 12 of the drawings becomes a vocal utterance in the vocal sound analyzer through a microphone 1210 introduced, and by a microphone amplifier 1211 for signal amplification, or from a recorded input through a ribbon input connector 1212 to input a pre-written vocal utterance. An input level control 1213 sets the voice signal level to the filter driver amplifier 1214 one. The filter driver amplifier 1214 amplifies the signal and sends the signal to the VU meter 1215 to measure the correct operating signal level.

The rate of flow per second and the number of passes per sample is determined by the operator with the sweep rate and sample time switches 1216 regulated or controlled. The operator starts a scan with the scan start switch and the stop override 1217 , The override feature allows the operator to manually override the set scan time and stop sampling to prevent contaminating a sample with unexpected sound disturbances, including simultaneous speakers. This switch also connects and disconnects the microprocessor power supply with 110 volt standard input electrical pins.

The output of the filter driver amplifier 1214 is also available to a commercially available microprocessor controlled filter bank and digitizer 1218 applied, which segments the electrical signal into 1/3 octave ranges over the audible frequency range for the organism which is sampled and digitizes the voltage output from each filter. In a specific work system, 25 1/3 octave filters of an Eventide spectrum analyzer range from 63 Hz to 16,000 Hz filter center frequencies. Further, an AKAI microphone and tape recorder with built-in amplifier was used as the input to the filter bank and digitizer 1218 used. The number of passes per second the filter bank uses is about ten passes per second. Other microprocessor-based filter banks and digitizers can operate at different speeds.

Any of various commercially available microprocessors suitable for the above Filterbank and the digitizer to regulate or control.

As with any complex sound, an amplitude over the audible frequency range will not be constant or flat for a "time fraction" of 0.1 of a second, but will rather be peaks and valleys. The values representative of a frequency of the peaks of this sig Nalles, 1219 are made more accurate by finding the amplitude values on each side of the peaks, and setting the peaks to the adjacent filter value having the larger amplitude. This is done because, as is characteristic of adjacent 1/3 octave filters, energy at a given frequency overflows to some extent into adjacent filters depending on the cut-off qualities of the filters. To minimize this effect, the frequency of a peak filter is assumed to be the center frequency only when the two adjacent filters have amplitudes within 10 of their average. To guarantee discrete, equally spaced, small values for linearizing and normalizing the values representing the unequal frequency intervals, each of the twenty-five filters are given numbers 1-25 and these numbers are used for the remainder of the processing. In this way, the 3500 Hz difference between the filters 24 and 25 becomes a value of 1, which in turn is also equal to the 17 Hz difference between the first and second filters.

And more than five Sub-subdivisions of each filter number to prevent and thus continue to evaluate equal steps between each Sub- or sub-division of 1 to 25 filter numbers maintain these are subdivided into 0.2 steps and further assigned as follows. If the amplitude difference of the two adjacent filters to a top filter larger than Is 30 of their average, then is the number or number of the top filter, closer the point halfway to the next To be filter count than it is from the top filter. This would be the Filter number of a peak filter, say, filter number 6.0, to cause it to increase to 6.4 or to be reduced to 5.6 when the larger adjacent filter is one higher or lower frequency represents. All other filter values of peak filters will automatically be the Value of its filter number +0.2 and -0.2 given when the larger the adjacent filter amplitudes each have a higher or lower frequency represents.

The segmented and digitally represented or represented vocalization signal 1219 becomes after the above frequency correction 1220 compressed to save memory space by discarding all but six amplitude peaks. The inventor found that six tips were sufficient to absorb the style characteristics as long as the following characteristics were observed. At least one peak is near the fundamental frequency; exactly one peak is allowed between the range of the fundamental frequency and the peak amplitude frequency, where the next one is conserved to the maximum and maximum peak, respectively; and the first two peaks above the maximum peak are stored plus the peak closest to the 16,000 Hz end or the 25 th filter, if above 8 kHz, for a total of six peaks stored and stored in the microprocessor memory. This will guarantee that the maximum peak is always the third peak stored in memory, and that the sixth stored peak can be used for high frequency analysis and that the first is the lowest and next to the fundamental.

Following the compression of the signal, an amplitude value of a complete band, the filter count and amplitude value of six peaks, and each of these thirteen values for 10 samples for a 10 second sample (FIG. 1300 Values), 1221 from 12 To include, the sound element assembly begins.

To arrive at vocal style "quality" elements, this system uses relationships between the low set and higher set of frequencies in the vocal utterance. On the other hand, the speech style elements are determined by a combination of measurements relating to the pattern of voice energy appearances, such as pauses and decay rates. These vocal style "quality" elements emerge from spectrum analysis, 13 . 1330 . 1331 , and 1332 on. The language style elements emerge from other four analysis functions, as in 12 . 1233 . 1234 . 1235 , and 1236 and Table 6 is shown.

The stored voice style quality analysis elements are referred to and derived as: (1) the spectrum "distribution" - the sample average of the spacing in filter numbers between the average of the peak filter numbers above and the average of the peak filter numbers below the maximum peak, for each pass; 13 . 1330 ; (2) the energy "balance" of the spectrum - the mean for a sample of all ratios of the passage of the sum of the amplitudes of those peaks above the sum of the amplitudes below the maximum peak, 1331 ; (3) the "flatness" spectrum envelope - the arithmetic mean for each of two sets of ratios for each sample - the average amplitude ratios of these peaks above (high) to the maximum peak, and from those below (low) the maximum peak to peak Maximum peak for each pass, 1332 ,

The language style elements which are stored are each designated and derived: (1) Spek trumvariability - the six average values of an utterance sample, the numerical differences between each filter number of a peak, on one pass, to each corresponding filter number of a peak on the next pass, and also the six amplitude value differences for those six Peaks and also includes the full spectrum amplitude differences for each pass to a sample of 13 means, 1333 to create; (2) pause ratio analysis - the ratio of the number of passes in the sample at which the complete energy amplitude values were paused (below two units of the amplitude value) to the number that had sound energy (greater than one unit of the value), 1334 ; (3) syllabic change approximation - the ratio of the number of passes at which the third peak changed the number value by more than 0.4 to the number of passes that had sound during the scan, 1335 ; (4) and, high-frequency analysis - the ratio of the number of passes for the sample at which the sixth peak had an amplitude value to the total number of passes, 1336 ,

sound styles are divided into the seven dimensions as shown in the table 6 is shown. These were the ones that were the most sensitive for one associated set of seven perceptual or recognition dimensions to be listed in Table 7.

The method of relating the sound style elements to the voice, speech, and perceptual dimensions of the output, 12 . 1228 , is done by equations that determine each dimension as a function of selected sound style elements, 13 . 1330 to 1336 , Table 6 relates the language style elements, 1333 to 1336 from 13 , on the language style dimensions.

• ##STR1##
• DS1 = Variabilität monoton
• DS2 = abgehackt sanft bzw. glatt
• DS3 = Stakkato aufrechterhalten
• DS4 = Anstieg sanft
• D55 = Affektivitätsregelung bzw. -steuerung
• (2) Nr. 1 bis 6 = Spitzenfilterunterschiede 1-6, und Amp1 bis 6 = Spitzenamplitudendifferenzen bzw. -unterschiede 1-6.
• Amp7 = Volle Bandpaßamplitudendifferenzen.

Tabelle 7

• ##STR2##
• DP1 = Eco-Struktur hoch-niedrig;
• DP2 = Invariantempfindlichkeit hoch-niedrig;
• DP3 = anders-selbst;
• DP4 = sensorisch-intern;
• DP5 = Haß-Liebe;
• DP6 Abhängigkeit-Unabhängigkeit;
• DP7 = emotionell-physisch.
• (2) Nr. 1 bis 6 = Spitzenfilterdifferenzen 1-6; Amp1 bis 6 = Spitzenamplitudendifferenzen 1-6; und Amp7 vollständige Bandpaßamplitudendifferenzen.

Table 7 shows the relationship between seven perceptual dimension and the sound style elements, 1330 to 1336 Again, the purpose is to have an optional input coefficient array which includes zeros to allow the device operator to make changes in these coefficients for research purposes, 1222 . 1223 to switch or enter. The smart operator may develop different perceptual dimensions, or even personality or cognitive dimensions, or factors (if he prefers this phraseology), all of which require different coefficients together. This is done by entering the desired set of coefficients and noting which dimension (s) 1226 ) he relates this. For example, the dimension of others-even of Table 7-need not be a desired dimension by a researcher who wants to replace them with a user perceptual dimension that he calls introvert-extroverted. By replacing the coefficient set for the set of others - even extroverted by experimental sentences until an acceptably high correlation between the selected combination of weighted sound style elements and its externally determined dimension is introverted - the explorer can then introvert - extrovert that slot for the new dimension by renaming them effectively. This can be done to the extent that the set of sound elements of this invention is sensitive to a user dimension of introvert-extrovert, and the set of coefficients of the researcher reflects the appropriate relationship. This will be possible with quite a few dimensions determined by a user to a useful extent, thereby allowing the system to function productively in a research environment where new perceptual dimensions related to sound style elements are explored, developed or evaluated. Table 6

• ## STR1 ##
• DS1 = variability monotone
• DS2 = choppy gentle or smooth
• DS3 = Stakkato maintained
• DS4 = rise gently
• D55 = affectivity control
• (2) Nos. 1 to 6 = peak filter differences 1-6, and Amp1 to 6 = peak amplitude differences 1-6.
• Amp7 = Full bandpass amplitude differences.

Table 7

• ## STR2 ##
• DP1 = Eco-structure high-low;
• DP2 = invariant sensitivity high-low;
• DP3 = different-self;
• DP4 = sensory-internal;
• DP5 = hate-love;
• DP6 dependence-independence;
• DP7 = emotional-physical.
• (2) Nos. 1 to 6 = peak filter differences 1-6; Amp1 to 6 = peak amplitude differences 1-6; and Amp7 full bandpass amplitude differences.

The primary results available to the user of this system are the dimension values, 1226 which is selectively controlled by a switch, 1227 available on a standard light display and also selectively for monitor, printer, modem and other standard output devices, 1228 to be displayed. These can be used to determine how close the subject's voice is to any or all of the sound or perceptual dimensions of the built-in or published or personally developed rules or standards, which can then be used to communicate with one another Help improve emotion recognition.

In In another exemplary arrangement, biosignals generated by a user, used to help To determine emotions in the language of the user. The recognition rate of a speech recognition system is compensated by changes improved in the language of the user, which from factors such as For example, emotion, anxiety or fatigue or tiredness result. One of a statement of one User derived speech signal is through a preprocessor modified and provided to a speech recognition system to improve the recognition rate. The speech signal is based modified on a biosignal, representing the emotional state of the user indicating or indicative.

Detailed illustrated 14 a voice recognition system where voice signals from the microphone 1418 and biosignals from the biomonitor 1430 through a preprocessor 1432 be received or recorded. The signal from the biomonitor 1430 to the preprocessor 1432 is a bio-signal indicative of the impedance between two points on the surface of a user's skin. The biomotor 1430 measures the impedance using a contact 1436 , which is attached to one of the user's fingers, and a contact 1438 , which is attached to another finger of the user. A biomonitor, such as a biofeedback monitor sold by Radio Shack, which is a division of Tandy Corporation, under the tradename (MICRONATA.RTM.BIOFEEDBACK MONITOR) Model Number 63-664, may be used. It is also possible to fix the contacts at other positions on the user's skin. When the user becomes agitated or anxious, the impedance between the points decreases 1436 and 1438 The decrease is done by the monitor 1430 which generates a biosignal which is indicative of a decreased impedance. The preprocessor 1432 uses the biosignal from the biomotor 1430 to that from the microphone 1418 modifying the received speech signal, wherein the speech signal is modified to compensate for the changes in the user's speech due to changes resulting from factors such as fatigue or a change in the emotional state. For example, the preprocessor 1432 the pitch of the speech signal from the microphone 1418 lower when the biosignal from the biomonitor 1430 indicates that the user is in an excited state and the preprocessor 1432 can the pitch of the speech signal from the microphone 1418 increase when the biosignal from the biomonitor 1430 indicates that the user is in a less agitated condition, such as fatigue. The preprocessor 1432 then sets the modified voice signal of the audio card 1416 available in a conventional manner. For purposes such as initialization or calibration, the preprocessor may 1432 with the PC 1410 communicate using an interface, such as an RS232 interface. The user 1434 can with the preprocessor 1432 by watching the ad 1412 and by entering commands using the keyboard 1414 or membrane keyboard 1439 or communicate with a mouse.

It is also possible to use the biosignal to control the speech signal by controlling the gain and / or frequency response of the microphone 1418 preprocess. The gain or gain of the microphone can be increased or decreased in response to the biosignal. The biosignal can also be used to change the frequency response of the microphone. For example, if the microphone 1418 an ATM71 model, available from AUDIO-TECHNICA US, Inc., is to use the biosignal to switch between a relatively flat response and an unrolled response, with the unrolled response providing less gain for low frequency speech signals ,

If the biomonitor 1430 the above-mentioned monitor, available from Radio Shack, is the biosignal in the form of a series of ramp-like signals, each ramp lasting about 0.2 ms. 15 illustrates the biosignal where a series of ramp-like signals 1542 is separated by a time T. The proportion or amount of time T between the ramps 1542 refers to the impedance between the points 1438 and 1436 , When the user is in a more excited state, the impedance between the points becomes 1438 and 1436 decreases and the time T is reduced. When the user is in a less-excited state, the impedance between the points becomes 1438 and 1436 increased and the time T is increased or increased.

The Form of a biosignal from a biomonitor can take other forms as a series of ramp-like Be signals. For example, the biosignal may be an analog signal which, in the periodicity, Amplitude and / or frequency varies based on measurements, which made by the biomonitor, or it can be a digital one Value based on conditions measured by the biomonitor.

The biomonitor 1430 includes the circuit of 16 which generates the biosignal, which is the impedance between the points 1438 and 1436 displays. The circuit consists of two sections. The first section is used to measure the impedance between the contacts 1438 and 1436 and the second section acts as an oscillator to produce a series of ramp signals at the output connector 1648 where the frequency of the oscillation is controlled by the first section.

The first section controls the collector current I _{c, Q1} and the voltage V _{c, Q1 of} the transistor Q1 based on the impedance between the contacts 1438 and 1436 , In this embodiment, the impedance sensor is made 1650 simply from contacts 1438 and 1436 which are or are positioned on the skin of the speaker. As the impedance between the contacts 1438 and 1436 relatively slow compared to the oscillation frequency of the section 2 changes, the collector current I _{c, Q1} and the voltage V _{c, Q1 are} virtual or nearly constant, as far as the section 2 is affected. The capacitor C3 further stabilizes these currents and voltages.

The section 2 acts as an oscillator. The reactive components, L1 and C1, turn transistor Q3 on and off to produce oscillation. When the power is first turned on, I _{c, Q1} turns on Q2 by pulling base current I _{b, Q2} . Similarly, I _{c, Q2} turns on transistor Q3 by providing a base current I _{b, Q3} . Initially there is no current through the inductor or the inductance L1. When Q3 is on, the voltage Vcc is applied less than a small saturated transistor voltage V _{c, Q3} across L1. As a result, the current I _{L1 increases} in accordance with FIG

As the current I _L1 increases, the current I _c1 through the capacitor C1 increases. Increasing the current I _c1 reduces the base current I _{B, Q2} from the transistor Q2, since the current I _{c, Q1 is} virtually constant. This in turn reduces the currents I _{c, Q 2} , I _{b, Q 3} and I _{c, Q 3} . As a result, more of the current I _L1 passes through the capacitor C1 and further reduces the current I _{c, Q3} . This feedback causes the transistor Q3 to be turned off. Finally, the capacitor C1 is fully charged and the currents I _L1 and I _c1 drop to zero, allowing the current I _{c, Q1} again to pull the base current I _b, Q2 and turn on the transistors Q2 and Q3, which is the oscillation cycle starts again.

The current I _{c, Q1} , which depends on the impedance between the contacts 1438 and 1436 depends regulates or controls the frequency of the load ratio or duty cycle of the output signal. When the impedance between the points 1438 and 1436 decreases, the time T between the ramp signals decreases, and when the impedance between the points 1438 and 1436 increases, the time T between the ramp signals increases.

The circuit is powered by a three volt battery source 1662 driven, which with the circuit via the switch 1664 connected is. Also included is a variable resistor 1666 , which is used to set an operating point for the circuit. It is desirable to use the variable resistor 1666 to a position which is approximately in the middle of its adjustment range. The circuit then varies from this operating point, as described earlier, based on the impedance between the points 1438 and 1436 , The circuit also includes a switch 1668 and a speaker 1670 , If a mating connector is not in the connector 1648 is inserted, the switch provides 1668 the output of the circuit on the speaker 1670 rather than at the connector 1648 to disposal.

17 is a block diagram of the preprocessor 1432 , An analog-to-digital converter or converter (A / D) 1780 receives a voice or utterance signal from the microphone 1418 , and an analog-to-digital converter (A / D) 1782 receives a biosignal from the biomonitor 1430 , The signal from the A / D 1782 becomes a microprocessor 1784 made available. The microprocessor 1784 monitors the signal from the A / D 1782 to determine what action to take by the Digital Signal Processor Device (DSP). 1786 should be made. The microprocessor 1784 uses a memory 1788 for a program storage and for clipboard operations. The microprocessor 1784 communicates with the PC 1410 using an RS232 interface. The software for controlling or controlling the interface between the PC 1410 and the microprocessor 1784 can on the pc 1410 in a multi-application environment using a software package, such as a program sold under the trade name (WINDOWS) by Microsoft Corporation. The output from the DSP 1786 becomes an analog signal through a digital to analogue converter 1790 reconverted. After the DSP 1786 the signal from the A / D 1780 modified, as by the microprocessor 1784 is commanded, the output of the D / A converter 1790 to the audio card 1416 sent. The microprocessor 1784 may be one of widely available microprocessors such as the microprocessors available from Intel Corporation and the DSP 1786 can be one of the widely available digital signal processing chips used by companies such as the TMS320CXX series of Texas Instruments devices.

It is possible to use the biomonitor 1430 and preprocessor 1432 to position on a single card, which is in a blank card slot in the PC 1410 is used. It is also possible to use the functions of the microprocessor 1784 and the digital signal processor 1786 using the PC 1410 instead of performing specialized hardware.

The microprocessor 1784 monitors the biosignal from the A / D 1782 to determine what action to take by the DSP 1786 should be made. If the signal from the A / D 1782 indicates that the user is in a more excited state, the microprocessor shows 1784 the DSP 1786 that it receives the signal from the A / D 1780 should process, so that the pitch of the speech signal is reduced. When the biosignal from the A / D 1782 indicates that the user is in a less agitated or tired state, the microprocessor instructs 1784 the DSP 1786 to increase the pitch of the speech signal.

The DSP 1786 modifies the pitch of the speech signal by generating a speech model. The DSP then uses the model to restore the speech signal at a modified pitch. The language model is generated using one of the linear predictive coding techniques that are well known in the art. One such technique is disclosed in an application book by Analog Device, Inc., entitled "Digital Signal Processing Applications Using the ADSP 2100 Family, pages 355-372, published by Prentice-Hall, Englewood Cliffs, NJ, 1992. This technique involves modeling the speech signal as a finite impulse response (FIR) filter with time-varying coefficients, the filter being a train Then the time T between the pulses is a measure of the pitch or fundamental frequency The time-variant coefficients can be calculated using a technique such as the Levinson-Durbin recursion described in the above-cited publication by Analog Device. Inc. A time T between the pulses that form the train of pulses that excite the filter can be calculated using an algorithm such as the SIFT (Simplified Inverse Filter Following) algorithm of John D. Markel is disclosed in "The SIFT Algorithm for Fundamental Frequency Estimation" by John D. Markel, IEEE Transactions on Audio and Electroacoustics, Vol. AU-20, No. 5, December 1972. The DSP 1786 modifies the pitch or fundamental frequency of the speech signal by changing the time T between the pulses when it energizes the FIR filter to restore the speech signal. For example, the pitch can be increased by 1% by reducing the time T between the pulses by 1%.

It should be noted that the Speech signal modified in other ways than pitch changes can be. For example, you can Pitch, Amplitude, frequency and / or signal spectrum to be modified. One Section of the signal spectrum or the entire spectrum may be attenuated or be strengthened.

It is possible, too, to monitor other biosignals as a signal, which is for the impedance between two points on a user's skin indicating. Signals for an autonomous activity can indicate be used as biosignals. Signals indicative of autonomic activity such as blood pressure, heart rate, brainwave or others electrical activity, pupil size, skin temperature, Transparency or reflectivity a certain electromagnetic wavelength, or other signals that for the indicating the user's emotional state can be used become.

18 illustrates pitch modification curves which the microprocessor 1784 used to the DSP 1786 to instruct the pitch of the speech signal based on the time period T associated with the Bi osignal is associated, change. The horizontal axis 1802 shows the time period T between the ramps 1442 of the biosignal on and the vertical axis 1804 indicates the percentage change in pitch caused by the DSP 1786 is introduced.

19 FIG. 11 illustrates a flow chart of the instructions issued by the microprocessor 1784 be executed to a in 18 to illustrate the illustrated operating line or curve. After an initialization becomes step 1930 executed to build a line that is co-linear with the axis 1802 is. This line indicates that a zero pitch change is introduced for all values of T from the biosignal. After the step 1930 becomes a decision step 1932 running where the microprocessor 1784 determines if a modification command or command from the keyboard 1414 or the membrane keyboard 1439 was received. If no modification command has been received, the microprocessor waits 1784 in a loop to a modification command. When a modification command is received, a step becomes 1934 to determine the value of T = T _ref1 which will be used to establish a new reference point Ref1. The value T _ref1 is equal to the current value of T obtained from the biosignal. For example, T _ref1 can be equal to 0.6 ms. After a determination of the value T _ref1 , the microprocessor performs 1784 one step 1938 which prompts the user to make an utterance, so that a pitch sample in step 1940 can be removed. It is desirable to obtain a pitch sample because the pitch sample is used as a basis for the percent changes in pitch along the axis 1804 is displayed. In step 1942 instructs the microprocessor 1784 the DSP 1786 to increase the pitch of the speech signal by an amount equal to the current pitch change associated with point Ref1 plus an increase of five percent, however, smaller or larger increments may be used. (At this point, the pitch change associated with point Ref1 is zero.) See or recall step 1930 .) In step 1944 demands the microprocessor 1784 the user to perform a recognition test by speaking various commands to the speech recognition system to determine if an acceptable recognition rate has been achieved. When the user completes the test, the user can stop the test to the microprocessor 1784 Show by using a command, such as "End", using the keyboard 1414 or membrane keyboard 1439 is entered.

After executing the step 1944 leads the microprocessor 1784 one step 1946 in which he is the DSP 1786 instructs to reduce the pitch of the incoming speech signal by the pitch change associated with point Ref1, minus a five percent reduction; however smaller or larger amounts or shares can be used. (Notice that the pitch change associated with point Ref1 becomes zero as a result of the step 1930 is). In step 1948 demands the microprocessor 1784 prompt the user to perform another speech recognition test and enter an "end" command when the test is complete. In step 1950 demands the microprocessor 1784 Ask the user to vote for the first or second test to indicate which test had better detection capability. In step 1952 the results of the user's choice are used between steps 1954 and 1956 select. If the test 1 was rated as the best, the step becomes 1956 and the new percentage change associated with the point Ref1 is set equal to the previous value of the point Ref1 plus five percent, or the increment, which in step 1942 has been used. If the test 2 is rated as the best, the step becomes 1954 and the new percent change value associated with Ref1 is set equal to the old value of Ref1 minus five percent, or the decrease made in step 1946 has been used. Determining a percentage _change associated with T = T _ref1 establishes a new reference point Ref1. For example, if the test 1 was rated as the best, the point Ref1 is at the point 1858 in 18 arranged. After building the position of the point 1858 , which is the newly built Ref1, is or will be the line 1860 in step 1962 built. The line 1860 is the initial pitch modification line which is used to calculate the pitch changes for different values of T from the biosignal. Initially, this line may be given a slope of, for example, plus five percent per millisecond; however, other slopes may be used.

After building this initial modification line, the microprocessor goes 1784 in a holding pattern where steps 1964 and 1966 be executed. In step 1964 checks the microprocessor 1784 after a modification command or command, and in step 1966 he checks for a shutdown command. If a modification command in step 1964 is not received, the processor checks after the shutdown command in step 1966 , If a shutdown command is not received, the microprocessor will return to step 1964 back, and when a shutdown command is received, the microprocessor steps 1930 which equates the pitch change with zero for all values of T from the biosignal. The processor remains in this loop for checking for modify and shutdown commands, until the user becomes dissatisfied with the recognition rate resulting from the pre-processing of the speech signal using the curve 1860 results.

When in step 1964 a modification command is received, becomes a step 1968 executed. In step 1968 the value of T is determined to check whether the value of T is equal to or nearly equal to the value of T _{ref1 of} the point Ref1. If the value of T matches Ref1, the step becomes 1942 executed. If the value of T does not match Ref1, the step becomes 1970 executed. In step 1970 the value of T _{ref2 is established} for a new reference point Ref2. For purposes of an illustrative example, we will assume that T _ref2 = 1.1 ms. With reference to 18 this builds the point Ref2 as a point 1872 on the line 1860 , In step 1974 instructs the microprocessor 1784 the DSP 1786 The pitch change associated with point Ref2 may increase by plus 2.5 percent (other percentages may be used). (Other percentages can be used). In step 1976 the user is prompted to perform a recognition test and to enter the "end" command or command at completion. In step 1978 instructs the microprocessor 1784 the DSP 1786 to decrease the pitch of the speech signal by an amount equal to a pitch change associated with Ref2 minus 2.5 percent. In step 1980 the user is again prompted to perform a recognition test and to enter an "end" command upon termination. In step 1982 the user is asked to indicate if the first or second test had the most desirable results. In step 1984 decides the microprocessor 1784 , one step 1986 if Test 1 was rated as the best, and a step 1988 if Test 2 was rated as the best. In step 1986 represents the microprocessor 1784 the percent change associated with point Ref2 is at the previous value associated with Ref2 plus 2.5 percent or the increase generated in step 1974 has been used. In step 1988 For example, the percentage change associated with Ref2 is set equal to the earlier value associated with Ref2 minus 2.5 percent or the decrease that was determined in step 1978 has been used. After completing the steps 1986 or 1988 becomes a step 1990 executed. In step 1990 is a new pitch modification line built. The new line uses the point associated with Ref1 and the new point associated with Ref2. For example, assuming that the user is testing 1 in step 1984 the new point associated with Ref2 has the dot 1892 from 18 , The new pitch conversion line is now the line 1898 which through the points 1892 and 1858 passes. After executing the step 1990 the microprocessor returns 1684 to the one with the steps 1964 and 1966 associated loop function back.

It should be noted that a linear modification line was used; however, it is possible to use non-linear modification lines. This can be done by the points 1858 and 196 used to be an increase for a line to the right of the point 1858 and by placing another reference point to the left of the point 1858 is used to build a slope for a line that is to the left of the point 1858 extends. It is also possible to arrange positive and negative limits on the maximum percentage pitch change. As the pitch modification line approaches these limits, they may approach it asymptotically or simply abruptly change at the point of contact with the boundary.

It is also possible to use a fixed modification curve, such as a curve 1800 , and then the variable resistor 1666 until an acceptable recognition rate is achieved.

Voice or Voice Notification System

20 FIG. 10 illustrates a system that handles voice messages based on emotion characteristics of the voice messages. In process 2000 A plurality of voice messages transmitted over a telecommunication network are received. In process 2002 For example, the voice messages are stored on a storage medium such as the above-mentioned tape recorder or a hard disk. An emotion associated with the voice signals of the voice messages is being processed 2004 certainly. The emotion may be determined by any of the methods set forth above.

The voice messages are in process or function 2006 organized based on the particular emotion. For example, messages in which the voice indicates negative emotions, such as sadness, anger or anxiety, may be grouped together in a mailbox or mailbox and / or database. Access to the organized voice messages is in progress 2008 allowed.

The voice messages can follow a telephone call. Optionally, the voice messages ei be organized together with a similar emotion. Also optionally, the voice messages may be organized in real-time immediately upon receipt via the telecommunications network. Preferably, a manner in which the voice messages are organized is identified to facilitate access to the organized voice messages. Also preferably, the emotion is determined by extracting at least one feature from speech signals, as previously discussed.

In In an exemplary arrangement, pitch and LPC parameters (and usually also other excitation information) for transmission and / or storage encoded, and are decoded to be a close replica of the original one Voice input available to deliver.

wobei u_k das LPC-Restsignal ist. Das heißt, u_k repräsentiert die verbleibende bzw. Restinformation in dem eingegebenen bzw. Eingabesprachsignal, welches nicht durch das LPC-Modell vorhergesagt ist. Es soll beachtet werden, daß nur N ältere bzw. frühere Signale zur Vorhersage verwendet werden. Die Modellreihenfolge (typischerweise etwa 10) kann erhöht werden, um eine bessere Voraussage zu ergeben, jedoch wird etwas Information immer in dem Restsignal u_k für jede normale Sprachmodellierungsanwendung verbleiben.The present system is particularly related to linear predictive coding (LPC) systems for analyzing or encoding analog speech signals (and methods therefor). In LPC modeling, generally, each sample in a series of samples (in the simplified model) is modeled as a linear combination of previous samples, plus an excitation function:

where u _{k is} the LPC residual signal. That is, u _k represents the residual information in the input speech signal which is not predicted by the LPC model. It should be noted that only N older or earlier signals are used for prediction. The model order (typically about 10) can be increased to give better prediction, but some information will always remain in the residual signal u _k for any normal speech modeling application.

Within The general framework of LPC modeling can be many special implementations of a voice analysis are selected. In many of these, it is necessary to change the pitch of the input speech signal to determine. That is, in addition to the formant frequencies, which actually coincide with resonances of the vocal tract, The human voice also includes a pitch given by the speaker is modulated, which coincides with the frequency at which the Larynx modulates the airflow. That is, the human voice can be regarded as an excitation function, which corresponds to an acoustic passive filter is applied, and the excitation function will generally appear in the LPC residual function while the Characteristics of the passive acoustic filter (i.e. the resonance characteristics of the mouth, nasal cavity, thorax, etc.) will be shaped by the LPC parameters. It should be noted be that while voiceless Language the excitation function does not have a well-defined pitch, but instead as broadband, white noise or pink noise is modeled.

A appraisal the pitch period is not complete trivial. Among the problems is the fact that the first Formant often will occur at a frequency near that of the pitch. For this Reason the sound level estimation often becomes performed the LPC residual signal, since the LPC estimation process indeed Vocal tach resonances unfolded from the excitation information, so that the residual signal relative less of the vocal tract (formants) and relatively more of the Excitation information (pitch) includes. However, such residue-based pitch estimation techniques their own difficulties. The LPC model itself usually becomes introduce high-frequency noise into the residual signal, and sections from this high-frequency noise can have a higher spectral density as the actual Pitch, which should be detected. One solution to this difficulty is simple, to filter the residual signal at about 1000 Hz lowpass. This removed the high-frequency noise, but also removes the legitimized High-frequency energy, which in the unvoiced areas of the language is present, and makes the residual signal almost useless for voiced Decisions.

One Main criterion in voice messaging applications is the quality of the reproduced Language. Prior art systems had in this regard many difficulties. In particular, many of these difficulties relate to problems of accurately detecting the pitch and voicing of the input speech signal.

It is typically very easy to incorrectly estimate a pitch period at twice or half its value. For example, when correlation methods are used, a good correlation at a period P guarantees a good correlation at a period 2P, and also means that the signal is more likely to show a good correlation at a period P / 2. However, such doubling and halving errors produce a very annoying reduction in voice quality. example for example, an erroneous halving of the pitch period will tend to produce a squeaky voice, and an erroneous doubling of the pitch period will tend to produce a rough voice. Moreover, it is likely that doubling or halving a pitch period will occur intermittently, so that the synthesized voice will tend to crack or scratch intermittently.

Preferred arrangements use an adaptive filter to filter the residual signal. By using a time varying filter having a single pole at the first reflection coefficient (k ₁ of the speech input), the high frequency noise is removed from the voiced periods of the speech, but the high frequency information is retained in the unvoiced speech periods. The adaptively filtered residual signal is then used as the input for the pitch decision.

It is necessary, the high-frequency or high-frequency information in the to withhold voiceless speech periods or to maintain better voicing / voicelessness decisions to allow. That is, the "voiceless" voicing decision is usually done when no strong pitch is found is, that is, when no correlation delay of the residual signal high normalized correlation value. However, if just a low-pass filtered Section of the residual signal during voiceless speech periods is tested, this partial or sub-segment of the residual signal have spurious correlations. That is, the danger is that the truncated residual signal, which passes through the fixed low-pass filter produced according to the state of the art, does not contain enough data, to be reliable to show that no Correlation during voiceless periods, and the additional, by the high-frequency Energy of unvoiced periods provided bandwidth necessary is to be reliable the spurious correlation delays ruled out which could otherwise be found.

A Improvement in pitch and voicing decisions are particularly critical for voice messaging systems, is also for other applications desirable. For example, a word recognition device, which pitch information involved Naturally a good pitch estimation method require. In similar Way becomes a pitch information sometimes used for speaker checking, in particular over one Telephone line, where a high-frequency information is partially lost is. About that would be out for future wide-area detection systems it desirable able to be to consider the syntactic information provided by the pitch is specified. In similar Way would be a good voicing analysis for some advanced speech recognition systems, e.g. Speech-to-text systems desirable.

The first reflection coefficient k ₁ is approximately related to the high / low frequency energy ratio and a signal. See RJ McAulay, "Designing a Robust Maximum Likelihood Pitch Estimator for Speech and Additional Noise", Technical Note, 1979-28, Lincoln Labs, June 11, 1979. For k ₁ close to -1, there is more low frequency energy in the signal than high frequency energy and vice versa close to 1. Thus, for k ₁ is low pass filtered by the use of k ₁ to determine the pole of a 1-pole deemphasis filter, the residual signal in the voiced speech periods and is high pass filtered in the unvoiced speech periods. This means that the formant frequencies are excluded from calculation of the pitch during the voiced periods while maintaining the necessary high-bandwidth information in the unvoiced periods for accurate detection of the fact that there is no pitch correlation.

Preferably a post-processing, dynamic programming technique is used, and not only an optimal pitch value, but also a to provide optimal voicing decision. That is, both pitch as well as voicing are tracked from frame to frame and one cumulative disadvantage for a sequence of frame pitches / voicing decisions is for accumulated various tracks to find the track which optimal pitch and voicing decisions. The cumulative disadvantage is obtained by introducing a frame error from a frame to the next goes. The frame error preferably not only disadvantages large deviations in the pitch period from frame to frame, but also penalizes pitch hypotheses, which have a relatively poor correlation "goodness" value, and also penalizes changes in the voicing decision, if the spectrum is relatively unchanged from Frame to frame is. This last feature of the frame transition error therefore enforces voucher transitions to the points of maximum spectral change.

The voice message system includes a voice input signal which is referred to as a time series s _i 1, which is provided on an LPC analysis block. The LPC analysis can be done by a wide variety of conventional techniques, but the final product is a set of LPC parameters and a residual signal u _i . The background of LPC analysis in general and various methods of extracting LPC parameters is found in numerous commonly known references, including Markel and Gray, Linear Prediction of Speed (US Pat. 1976 ) and Rabiner and Schafer, Digital Processing of Speed Signals ( 1978 ) and cited therein.

In the presently preferred arrangement, the analog speech waveform is sampled at a frequency of 8 KHz and with an accuracy of 16 bits to produce the input time series s _i . Of course, the system is not at all dependent on the sampling rate or the accuracy used and is applicable to speech sampled at any rate or with any degree of accuracy.

In the presently preferred arrangement, the set of LPC parameters which is used includes a plurality of reflection coefficients k _i , and a 10th-order LPC model is used (ie, only the reflection coefficients k ₁ to k ₁₀ are extracted). and higher order coefficients are not extracted). However, other model orders or other equivalent sets of LPC parameters may be used, as is known to those of skill in the art. For example, the LPC prediction _coefficients a _k may be used or the impulse response estimates e _k . However, the reflection coefficients k _{i are} most convenient.

In the present preferred arrangement, the reflection coefficients according to the Leroux-Gueguen method extracted, for example, in the IEEE Transactions on Acoustic, Speech and Signal Processing, page 257 (June 1977).

however could other algorithms that are good for those with experience in engineering are known, such as Durbin's are used to the coefficients to calculate.

A by-product of calculating the LPC parameters will typically be a residual signal u _k . However, if the parameters are calculated by a method that does not automatically output u _k as a by-product, the remainder can be found simply by using the LPC parameters to configure a finite impulse response digital filter which directly generates the residual series _uk is calculated from the input or input series _sk.

The residual signal time series u _k is now given by a very simple digital filtering process, which depends on the LPC parameters for the current frame. That is, the speech input signal s _k is a time series which has a value which may change once every sampling at a sampling rate of, for example, 8 KHz. However, the LPC parameters are normally recalculated only once every frame period at a frame rate of, for example, 100 Hz. The residual signal u _k also has a period equal to the sampling period. Thus, the digital filter, the value of which depends on the LPC parameters, is preferably not reset for each residual signal u _k . In the presently preferred arrangement, approximately 80 values in the residual signal time series u _k pass through the filter 14 before a new value of the LPC parameters is generated, and therefore is a new characteristic for the filter 14 implemented.

More specifically, the first reflection coefficient k _{1 is} extracted from the set of LPC parameters transmitted by the LPC analysis section 12 are provided. While the LPC parameters themselves are the reflection coefficients k ₁ , it is only necessary to look for the first reflection coefficient k ₁ . However, where other LPC parameters are used, transforming the parameters to produce the first-order reflection coefficient is typically extremely simple, for example, k 1 = a 1 / a 0

The System preferably uses the first reflection coefficient, to define 1-pole adaptive filter. However, the filter has to not a single-pole filter, but can be considered a more complex Filter configured to one or more poles or a or several zeros, some or all of which are adaptive can be varied.

It should also be noted that the adaptive filter characteristic of the adaptive filter need not be determined by the first reflection coefficient k ₁ . As is good in the art If there are many equivalent sets of LPC parameters, and the parameters in other LPC parameter sets can also provide desirable filter characteristics. Especially in any set of LPC parameters, the lowest order parameters are most likely to provide information about the coarse spectral shape. Thus, an adaptive filter could use a ₁ or e ₁ to define a pole, which may be a single or multiple pole, and may be used alone or in combination with other zeros and / or poles. Moreover, the pole (or zero), which is adaptively defined by an LPC parameter, does not have to coincide exactly with this parameter, but may be shifted in magnitude or phase.

Thus, the 1-pole adaptive filter filters the residual signal time series u _k to produce a filtered time series u ' _k . As discussed above, this filtered time series u ' _{k will have} significantly reduced its high-frequency energy during the voiced speech segments, but nearly the complete frequency bandwidth will be maintained throughout the unvoiced speech segments. This filtered residual signal u ' _k is then subjected to further processing to extract the pitch candidate and the voicing decision.

A wide variety of methods for extracting the pitch information consists of a residual signal and any of these can be used become. Many of these are generally discussed in the book mentioned above discussed by Markel and Gray.

wo u'_j das gefilterte Restsignal ist, k_min und k_max die Grenzen für die Korrelationsverzögerung k definieren, und m die Anzahl von Abtastungen in einer Rahmenperiode (80 in der bevorzugten Anordnung) ist und deshalb die Anzahl von zu korrelierenden Abtastungen definiert. Die Kandidatentonhöhenwerte sind bzw. werden durch die Verzögerungen k* definiert, bei welchem der Wert von C(k*) ein örtliches Maximum annimmt, und der skalare Wert von C(k) verwendet wird, um einen "Güte"-Wert für jeden Kandidaten k* zu definieren.In the presently preferred arrangement, the candidate pitch values are obtained by finding the peaks in the normalized correlation function of the filtered residual signal, defined as follows:

where u ' _{j is} the filtered residual signal, k _min and k _max define the boundaries for the correlation delay k, and m is the number of samples in a frame period (80 in the preferred arrangement) and therefore defines the number of samples to be correlated. The candidate pitch values are defined by the delays k * at which the value of C (k *) takes a local maximum and the scalar value of C (k) is used to obtain a "goodness" value for each candidate k * to define.

Optionally, a threshold value C _{min is introduced} on the quality measure C (k), and local maxima of C (k) that do not exceed the threshold C _min are ignored. If no k * exists for which C (k *) is greater than C _min , then the frame is necessarily unvoiced.

Alternatively, the quality threshold C _{min may be} omitted and the normalized autocorrelation function 1112 can be easily controlled to identify a given number of candidates having the best quality values, eg the 16 pitch period candidates k, which have the largest values of C (k).

In one arrangement, no threshold is imposed on or superimposed on the quality value C (k), and no voicing decision is made at this stage. Instead, the 16 pitch period candidates k * ₁ , k * ₂ , etc., along with the corresponding quality value (C (k * _i )) are reported for each individual. In the presently preferred arrangement, the voicing decision is not made at this stage, even if all of the C (k) values are extremely low, but the voicing decision is made in the subsequent dynamic programming step, discussed below.

In the presently preferred arrangement, a variable number of pitch candidates is identified according to a peaking algorithm. That is, the graph of "goodness" values C (k) compared with the candidate pitch period k is tracked. Each local maximum is identified as a possible peak. However, the presence of a peak at this identified local or local maximum is not confirmed until after that the function has dropped by a constant amount. This confirmed local maximum will then provide one of the pitch period candidates. After identifying each leading candidate in this way, the algorithm then searches for a valley. That is, each local minimum is identified as a possible valley, but is not acknowledged as a valley until thereafter the function has increased by a predetermined constant value. The valleys who does not honor or report separately, however, a confirmed valley will be needed after a confirmed peak is identified before a new peak. In the presently preferred embodiment, where the quality values are defined to be bounded by +1 or -1, the constant value required to confirm a peak or valley has been set to 0.2, but this can be largely changed. Thus, this stage provides a variable number of pitch candidates as output, from zero to 15.

In the present preferred arrangement is the set of pitch period candidates, which by the previous steps are provided, then provided to a dynamic programming algorithm. This dynamic programming algorithm then tracks both pitch and pitch also voicing decisions to make a pitch and Vocabulary decision for to provide every frame which is optimal in the context of his Neighbors is.

in view of the candidate pitch values and their quality values C (k) is now using dynamic programming to get an optimal pitch contour or outline to get an optimal voicing decision for each one Frame includes. Dynamic programming requires different ones Speech frames in a segment of speech are analyzed before the pitch and voicing for the first frame of the segment can be decided. At each Within the speech segment, each pitch candidate is withheld pitch candidates of the previous frame. Each withheld pitch candidate from the previous frame entails a cumulative or increasing disadvantage or deduction, and every comparison between a new pitch candidate and each of the retained pitch candidates also has a new distance or distance measure. Thus there is for everyone pitch candidates in the new frame a slight drawback or deduction, which a best match with one of the with or withheld pitch candidates of the previous frame. When calculating the smallest cumulative disadvantage for each new candidate became the candidate along with his cumulative disadvantage and a backward pointer to the best match retained in the previous frame. Thus, the backward pointers define a trajectory or state curve, which is a cumulative deduction as in the cumulative subtraction value of the last frame was recorded in the project rate. The optimal trajectory for each given frame is selected by selecting the trajectory with the receive minimal cumulative disadvantage. The unvoiced state is as a pitch candidate defined on every frame. The deduction or disadvantage function includes preferably voicing information so that the voicing decision a natural one Consequence of the dynamic programming strategy is.

In the present preferred arrangement is the dynamic programming strategy 16 wide and 6 deep. That is, 15 candidates (or less) plus the "voicelessness" decision (for Convenience as a zero pitch period) as possible pitch period identified on each frame, and all 16 candidates together with their quality values be for retained the 6 previous frames.

The pitch and voicing decisions are finally made only with respect to the oldest frame included in the dynamic programming algorithm. That is, the pitch and voicing decision would accept the candidate pitch at frame F _K -5 whose current trajectory cost was minimal. That is, out of the 16 (or fewer) trajectories ending in the most recent frame F _K , the candidate pitch in frame F _K , which has the lowest cumulative trajectory cost, identifies the optimal trajectory. This optimal trajectory is then traced back and used to make the pitch / voicing decision for frames F _K -5. It should be noted that no final decision is made about the pitch candidates in subsequent frames (F _k -4, etc.) because the optimal trajectory may no longer appear optimal after multiple frames are evaluated. Of course, as is well known to those of skill in the art and numerical optimization, a final decision in such a dynamic programming algorithm may alternatively be made at other times, eg, in the next to last frame held in the buffer becomes. In addition, the width and depth of the buffer can be varied widely. For example, up to 64 pitch candidates can be rated, or as few as two; the buffer could hold back as little as a previous frame, or as many as 16 previous frames or more, and other modifications and alterations can be made as will be appreciated by those of skill in the art. The dynamic programming algorithm is defined by the transition error between a pitch period candidate in one frame and another pitch period candidate in the subsequent frame. In the presently preferred arrangement, this transient error is defined as the sum of three parts: an error E _p on reason of pitch deviations, an error E _s due to pitch candidates having a low "goodness" value, and an error E _t due to the voicing transition.

wenn beide Rahmen stimmhaft sind, und E_P = B_P mal D_N andernfalls; wo tau die Kandidatentonhöhenperiode des gegenwärtigen Rahmens ist, tau_P eine zurückgehaltene Tonhöhenperiode des vorigen Rahmens, in bezug auf welchen der Übergangsfehler berechnet wird, ist, und B_P, A_D und D_N Konstante sind. Es soll beobachtet werden, daß die Minimumfunktion eine Vorkehrung zur Tonhöhenperiodenverdopplung und Tonhöhenperiodenhalbierung beinhaltet. Diese Vorkehrung ist nicht unbedingt notwendig, wird jedoch als vorteilhaft betrachtet. Natürlich könnte optional eine ähnliche Vorkehrung zur Tonhöhenperiodenverdreifachung beinhaltet sein, usw.The pitch deviation error E _p is a function of the current pitch period and the previous pitch period, indicated by:

if both frames are voiced, and E _P = B _P times D _N otherwise; where tau is the candidate pitch period of the current frame, tau _{P is} a retained pitch period of the previous frame with respect to which the transition error is calculated, and B _P , A _D and D _{N are} constants. It should be noted that the minimum function includes provision for pitch-period doubling and pitch-period bisecting. This provision is not strictly necessary but is considered advantageous. Of course, an optional provision for pitch tripling could optionally be included, etc.

The voicing state error E _S is a function of the "goodness" value C (k) of the currently considered frame pitch candidate. For the unvoiced candidate which is always included among the 16 or less pitch period candidates to be considered for each frame, the quality value C (k) is set equal to the maximum of C (k) for all the other 15 pitch period candidates in the same frame. The voicing state error E _S is given by E _S = B _S (R _v - C (tau) if the current candidate is voiced, and E _S = B _s (C (tau) - R _U ) otherwise where C (tau) is the "quality value" corresponding to the current pitch candidate tau, and B _S , R _V , and R _{U are} constants.

wo E die RMS-Energie bzw. -Effektivwertenergie des gegenwärtigen Rahmens ist, E_P die Energie des vorigen Rahmens ist, L(N) ist der N-te Logarithmus des Flächenverhältnisses des augenblicklichen Rahmens und L_P(N) N-te Logarithmus des Flächenverhältnisses des vorigen Rahmens ist. Das logarithmische Flächenverhältnis L(N) wird direkt aus dem N-ten Reflexionskoeffizienten k_N berechnet wie folgt:

The voicing transition error E _T is defined with respect to a spectral difference measure T. The spectral difference measure T, for each frame, generally defines how different its spectrum is from the spectrum of the receiving frame. Obviously, a number of definitions could be used for such a spectral difference measure defined in the presently preferred arrangement as follows:

where E is the RMS energy of the current frame, E _{P is} the energy of the previous frame, L (N) is the Nth logarithm of the area ratio of the current frame, and L _P (N) is the Nth logarithm of the current frame Area ratio of the previous frame is. The logarithmic area ratio L (N) is calculated directly from the Nth reflection coefficient k _N as follows:

The voicing transition error E _T is then defined as a function of the spectral difference measure T as follows:
If the current and previous frames are both unvoiced, or if both are voiced, E _T = 0 is set;
otherwise, E _T = G _T + A _T / T where T is the spectral difference measure of the spectral difference of the current frame. Again, the definition of voicing transition error could be widely varied. The key feature of the voicing transition error as defined herein is that whenever a voicing state change (voiced to unvoiced or voiceless to voiced) occurs, a penalty is found which is a decreasing function of the spectral difference between the two frames is. That is, a change in the voicing state is disfavored unless a considerable spectral change also occurs.

A Such definition of voicing transition error provides considerable Benefits because it reduces the processing time required is to deliver excellent voicing decisions.

The other errors E _s and E _p which make up the transient error in the presently preferred arrangement may also be variously defined. That is, the voicing state error may be defined in any manner which in the general pitch-period hypotheses that appear to fit the data in the current frame is preferable to those that are less suitable for the data. Similarly, the pitch deviation error E _{p may} be defined in any manner which generally coincides with or corresponds to changes in the pitch period. It is not necessary for the pitch deviation error to include a provision for doubling and halving, as defined herein, although such provision is desirable.

One Another optional feature is that when the pitch deviation error Arrangements for tracking pitch over doubling and halving includes, it desirable may be the pitch period values to double (or halve) along the optimal trajectory after the optimal trajectory has been identified, so to speak as far as possible consistent or uniform.

It should be noted that it not necessary, all of the three identified components of the transient error to use. For example, could the voicing state error should be omitted if some previous levels Pitch hypotheses excluded with a low "goodness" value, or if the pitch periods on the "goodness" value in a certain Were arranged such that the pitch periods, which have a higher quality value, would be preferred or otherwise. In similar Way you can other components in the transition error definition to be included as desired.

It should also be noted that the dynamic programming methods taught herein necessarily on pitch period candidates must be applied which have been extracted from an adaptively filtered residual signal, still on pitch period candidates, which were derived from the LPC residual signal, but on each Set of pitch period candidates can be applied which include pitch period candidates, directly from the original Input speech signal were extracted.

These Three errors are then summed to the total error between any one pitch candidates in the present Frame and any pitch candidate in the previous frame. As noted above became, these transitional errors then summed cumulatively, to cumulative penalty for each trajectory available in the dynamic programming algorithm put.

This dynamic programming method for simultaneous finding from both pitch as well as voicing is in itself novel, and does not have to only in combination with the currently preferred methods a finding of pitch period candidates be used. Any method for finding pitch candidates Can be used in combination with this novel dynamic programming algorithm become. No matter which method is used to pitch period candidates To find the candidates are simply as input to the dynamic Programming algorithm available posed.

Especially is while using a minicomputer and high-precision sampling are presently preferred This system is not economical for large volume applications. Consequently is expected from the preferred system in the future, an arrangement to be using a microcomputer based system, such as the TI Professional Computer.

This professional computer when using a microphone, speakers and a voice processing card is configured, including a TMS 320 numeric processing microprocessor and data converter, is sufficient hardware to implement the system.

Claims (8)

A method of detecting emotions in a voice by using statistics, comprising the steps of: (a) providing a database having statistics including statistics of human associations of speech parameters with emotions ( 600 ), wherein the statistics in the database include at least one of self-knowledge statistics and performance and performance confusion statistics; (b) receiving a voice signal ( 602 ); (c) extracting at least one feature of the speech signal ( 604 ); (d) comparing the extracted speech feature with the speech parameters in the database ( 606 ); (e) selecting an emotion from the database based on the comparison of the extracted speech feature with the speech parameters ( 608 ); and (f) issuing the chosen emotion ( 610 ). A computer program embodied on a computer readable medium for detecting emotion in a voice by using statistics, comprising: (a) a code segment providing a database having statistics including Statistics of human associations of speech parameters with emotions ( 600 ), the statistics in the database including at least one of self-knowledge statistics and performance and performance confusion statistics; (b) a code segment containing a speech signal ( 602 ) receives; (c) a code segment containing at least one feature of the speech signal ( 604 ) extracted; (d) a code segment representing the extracted speech feature with speech parameters in the database ( 606 ) compares; (e) a code segment representing an emotion from the database based on the comparison of the extracted speech feature with the speech parameters ( 608 ) chooses; and (f) a code segment representing the chosen emotion ( 610 ) when the program is run on a computer. A system for detecting emotion in a voice using statistics, comprising: (a) a logic for providing a database having statistics containing statistics of human associations of speech parameters with emotions ( 600 ), the statistics in the database including at least one of self-knowledge statistics and performance and performance confusion statistics; (b) a logic for receiving a speech signal ( 602 ); (c) a logic for extracting at least one feature of the speech signal ( 604 ); (d) a logic for comparing the extracted speech feature with the speech parameters in the database ( 606 ); (e) a logic for selecting an emotion from the database based on the comparison of the extracted speech feature with the speech parameters ( 608 ); and (f) a logic for outputting the chosen emotion ( 610 ). Invention according to one of the preceding claims, the database being probabilities of special language features includes, which are associated with an emotion. The invention of claim 4, wherein the selection of Emotion from the database analyzing the probabilities and a selection the most likely emotion based on the probabilities includes. The invention according to any one of the preceding claims, wherein the at least one extracted feature of the speech signal at least one from a slope of a fundamental frequency and a region of a first formant. The invention according to any one of the preceding claims, wherein the at least one extracted feature of the speech signal at least one of a maximum and maximum value of the energy, an area the energy, a maximum value of the first formant and an area of the second formant. The invention of any one of the preceding claims, wherein the at least one extracted feature of the speech signal is the maximum value of the fundamental frequency, the standard deviation of the fundamental frequency, the range of the fundamental frequency, the mean of the fundamental frequency, the mean of the bandwidth of the first formant, the mean of the bandwidth of the second formant, the standard deviation of the energy and the speech rate.